{"schema":"scienceos.atlas.feed-shard/1","feed":"trending","date":"2026-10-22","generatedAt":"2026-10-22T08:36:18Z","sources":["arxiv","openalex","huggingface"],"items":[{"paper":{"id":"arxiv:2610.26860","arxivId":"2610.26860","doi":"10.48550/arXiv.2610.26860","title":"OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory","abstract":"Multimodal large language models (MLLMs) have made rapid progress on visual reasoning, yet many tasks require evidence that is difficult to obtain through internal reasoning alone: our controlled study on 22 benchmarks shows that simply increasing the inference budget of the base model yields only modest gains. Such tasks demand grounded perception, external knowledge, and numerical measurements obtained through heterogeneous tools, whose outputs must be made interpretable to the model and retained across reasoning steps. We introduce OpenVLHarness, a multimodal reasoning harness organized around a capability layer and persistent multimodal memory. The capability layer combines task-level tool interfaces with evidence rendering, providing consistent access to heterogeneous backends and converting their outputs into model-readable observations. Persistent multimodal memory retains the corresponding images, structured data, and web references as addressable artifacts throughout each query. A coding interface executes generated programs over stored data, enabling computations that combine results from multiple tools. Together, these mechanisms support an iterative process of acquiring, inspecting, and composing multimodal evidence. Evaluated on 22 benchmarks with eight open and proprietary backbones, OpenVLHarness improves average scores by 9.7-12.6 points over matched no-tool baselines, including +11.2 with Qwen3-VL-8B, +9.7 with Kimi K3, and +9.9 with GPT-6 Astra. With GPT-6 Luna at medium reasoning effort, OpenVLHarness outperforms the base model, Codex, and Visual Sketchpad at maximum effort by 7.7-8.5 points while incurring at most 37% of their model-token cost.","authors":[{"name":"Yu Zhou","position":0},{"name":"Cheng-Fu Yang","position":1},{"name":"Rui Sun","position":2},{"name":"Di Wu","position":3},{"name":"Sibo Peng","position":4},{"name":"Mingyang Zhang","position":5},{"name":"Zhen Yang","position":6},{"name":"Ruiyi Zhang","position":7},{"name":"Nanyun Peng","authorId":"oa:A5030248499","position":8},{"name":"Zhe Gan","authorId":"oa:A5066666034","position":9},{"name":"Kai-Wei Chang","authorId":"oa:A5073201681","position":10}],"categories":["cs.CL","cs.AI","cs.CV","cs.LG","cs.MM"],"primaryCategory":"cs.CL","publishedAt":"2026-10-19T17:59:58Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26860"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26860"},{"kind":"code","url":"https://github.com/uclanlp/OpenVLHarness","label":"uclanlp/OpenVLHarness"},{"kind":"project","url":"https://openvlharness.github.io/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":104,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":167,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"uclanlp/OpenVLHarness"},"brief":{"text":"OpenVLHarness combines visual tools, persistent multimodal memory, and code execution to help models gather and use evidence across reasoning steps. Across 22 benchmarks, it raises average scores by 9.7–12.6 points over matched no-tool baselines, often at lower token cost than other maximum-effort approaches.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":1,"reason":"trending","score":0.8652},{"paper":{"id":"arxiv:2610.22471","arxivId":"2610.22471","doi":"10.48550/arXiv.2610.22471","title":"Has GPT-6-Astra Achieved Embodied Intelligence?","abstract":"GPT-6-Astra demonstrates a level of competence in robotic manipulation and navigation that challenges the conventional paradigm of building specialized models for individual tasks and embodiments. In this report, we ask not only how capable GPT-6-Astra is, but what its capabilities imply for the field. First, we probe the limits of standalone GPT-6-Astra. We conduct a large-scale evaluation across 12 representative benchmarks spanning tabletop manipulation, humanoids, mobile platforms, and dexterous hands. Under a unified protocol that allows in-context learning (ICL) from training-set demonstrations and substantially increases the number of execution steps, GPT-6-Astra achieves unprecedented performance (Fig. 1). Future research may focus on improving 1) the generality of ICL 2) cost and execution time, 3) physical-state verification, 4) recovery from irreversible failures, and 5) long-horizon preservation of progress. Second, we study the agentic harness for embodied control. We identify that close-loop coordination between a high-level \"System 2\" reasoner (i.e., GPT-6-Astra) and a low-level \"System 1\" controller (e.g., π0.5) is costly but effective. Third, we examine GPT-6-Astra on real-world deployment. Our experiments on 8 robot configurations reveal substantial gaps: 1) robot demonstrations required for ICL are expensive to collect, and 2) it is implausible to trade longer execution for higher success rates with physical constraints. Fourth, we envision the next generation of embodied benchmarks. We develop a prototype that 1) derives tasks from real-world needs, 2) provides no in-domain demonstrations, and 3) judges agents on a world clock that keeps running while they deliberate. Taken together, our results suggest that GPT-6-Astra has largely crossed the threshold of embodied competence, but not yet that of practical embodied intelligence.","authors":[{"name":"SenseNova","authorId":"oa:A5159035411","position":0},{"name":"Joint Research Team","authorId":"oa:A5159084667","position":1},{"name":"Zhongang Cai","authorId":"oa:A5054947207","position":2},{"name":"Ziang Cao","position":3},{"name":"Tailai Chen","position":4},{"name":"Han Chong Chua","authorId":"oa:A5159079748","position":5},{"name":"Linjun Dai","authorId":"oa:A5100962735","position":6},{"name":"Hanming Deng","position":7},{"name":"Hokin Deng","authorId":"oa:A5159093572","position":8},{"name":"Haiwen Diao","authorId":"oa:A5015049935","position":9},{"name":"Chenjing Ding","authorId":"oa:A5005400846","position":10},{"name":"Weichen Fan","authorId":"oa:A5027750483","position":11},{"name":"Xiao Fu","position":12},{"name":"Zerui Gong","authorId":"oa:A5134688170","position":13},{"name":"Chenyang Gu","position":14},{"name":"Zexin He","authorId":"oa:A5159068319","position":15},{"name":"Zhi Hou","position":16},{"name":"Bo Li","position":17},{"name":"Qinyue Li","authorId":"oa:A5053537085","position":18},{"name":"Quanzhou Li","authorId":"oa:A5113155097","position":19},{"name":"Xiaowan Li","authorId":"oa:A5159048050","position":20},{"name":"Zhichao Liao","position":21},{"name":"Dahua Lin","position":22},{"name":"Zilong Lin","authorId":"oa:A5080383166","position":23},{"name":"Bo Liu","position":24},{"name":"Chunxiao Liu","position":25},{"name":"Pengfei Liu","position":26},{"name":"Tianqi Liu","position":27},{"name":"Ziwei Liu","position":28},{"name":"Chen Change Loy","position":29},{"name":"Lewei Lu","position":30},{"name":"Jiapeng Luo","position":31},{"name":"Yuexiao Ma","authorId":"oa:A5158724314","position":32},{"name":"Liang Pan","position":33},{"name":"Xingang Pan","position":34},{"name":"Oscar Qian","position":35},{"name":"Ziyi Shang","position":36},{"name":"Yuyao Shao","position":37},{"name":"Dawei Su","position":38},{"name":"Dacheng Tao","authorId":"oa:A5101409301","position":39},{"name":"Qingyi Tao","position":40},{"name":"Quan Wang","position":41},{"name":"Ruisi Wang","position":42},{"name":"Xiaogang Wang","position":43},{"name":"Ziyi Wang","position":44},{"name":"Beichen Wen","position":45},{"name":"Donghao Wu","position":46},{"name":"Jiaxu Wu","position":47},{"name":"Silei Wu","position":48},{"name":"Xiaoming Wu","position":49},{"name":"Zhonghua Wu","position":50},{"name":"Hongcan Xiao","position":51},{"name":"Weiye Xu","position":52},{"name":"Atsushi Yamashita","position":53},{"name":"Lei Yang","position":54},{"name":"Zhe Yang","authorId":"oa:A5159080957","position":55},{"name":"Baiqiao Yin","position":56},{"name":"Wanqi Yin","position":57},{"name":"Hanfei Zhao","authorId":"oa:A5159042277","position":58},{"name":"Tianbo Zhong","position":59},{"name":"Long Zhuo","position":60},{"name":"Jing Zuo","position":61}],"categories":["cs.RO","cs.CV","cs.LG","cs.MM"],"primaryCategory":"cs.RO","publishedAt":"2026-10-15T18:00:00Z","updatedAt":"2026-10-19T17:20:19Z","version":2,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.22471"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.22471"},{"kind":"code","url":"https://github.com/OpenSenseNova/SenseNova-RoboRSI","label":"OpenSenseNova/SenseNova-RoboRSI"},{"kind":"project","url":"https://sensenova.ai/has-gpt-6-astra-achieved-embodied-intelligence"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":460,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":224,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"OpenSenseNova/SenseNova-RoboRSI"},"brief":{"text":"GPT-6-Astra achieves unprecedented results across 12 simulated robotics benchmarks, but real-world tests on eight robot configurations expose costly demonstrations and physical limits; it shows embodied competence, not yet practical embodied intelligence.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":2,"reason":"trending","score":0.8251},{"paper":{"id":"arxiv:2610.26825","arxivId":"2610.26825","doi":"10.48550/arXiv.2610.26825","title":"MemT-1.0: Towards Memory-Native Evolution","abstract":"Memory-native evolution requires making memory an intrinsic capability of foundation models, enabling them to continuously accumulate and utilize information. To this end, we introduce MemT-1.0, a new architecture of memory foundation models that internalize memory as part of model computation. Building on Metis, MemT-1.0 employs Gated DeltaNet-2 with a stable token-level memory update mechanism within each turn. We further introduce Mixture-of-Storage to scale native memory capacity. Following a deliberate backbone preparation, we develop a two-stage training pipeline consisting of continued pre-training and post-training, supported by 34B tokens of training data and efficient training and inference infrastructure. We release a model family comprising MemT-1.0-Nano and MemT-1.0-Flash, and conduct a comprehensive evaluation of memory and general capabilities. The results show that they outperform the parametric memory baselines on average and remain competitive with retrieval-augmented methods, while largely preserving general capabilities. At a history length of 128K tokens, MemT-1.0-Nano achieves approximately $1.86\\times$ end-to-end inference speedup and a $66\\times$ reduction in persistent storage relative to full context KV cache. By internalizing retrieved procedural experience, MemT-1.0 improves task success on agentic tasks, providing a practical step toward memory-native evolution.","authors":[{"name":"Zhiyu Li","position":0},{"name":"Zeyu Zhang","position":1},{"name":"Linyun Liu","authorId":"oa:A5088841352","position":2},{"name":"Yiwei Bu","position":3},{"name":"Chenyang Xi","authorId":"oa:A5058308218","position":4},{"name":"Xichong Zhang","position":5},{"name":"Yi Yu","position":6},{"name":"Yu Yu","position":7},{"name":"Zhichen Liu","position":8},{"name":"Yuxiao Yi","authorId":"oa:A5021663594","position":9},{"name":"Chunyu Li","position":10},{"name":"Haoran Liu","position":11},{"name":"Zhenkai Ding","authorId":"oa:A5134121373","position":12},{"name":"Le Zhou","position":13},{"name":"Juncheng Zhang","authorId":"oa:A5127157050","position":14},{"name":"Boyu Chen","position":15},{"name":"Zehao Lin","position":16},{"name":"Guorui Li","authorId":"oa:A5121238029","position":17},{"name":"Renyu Fu","authorId":"oa:A5133364288","position":18},{"name":"Ji Li","position":19},{"name":"Youyang Xia","authorId":"oa:A5010101173","position":20},{"name":"Jiajun Shen","position":21},{"name":"Harvey Li","authorId":"oa:A5135432167","position":22},{"name":"Jingsong Zhou","position":23},{"name":"Helan Hu","position":24},{"name":"Tong Shen","position":25},{"name":"Tianxiao Xu","position":26},{"name":"Chenye Han","position":27},{"name":"Weijie Li","position":28},{"name":"Zhongquan Zhou","position":29},{"name":"Hongming Xu","position":30},{"name":"Yuekeng Li","position":31},{"name":"Erenkai Zhu","position":32},{"name":"Peng Yu","position":33},{"name":"Wenhao Li","position":34},{"name":"Jingyi Kang","position":35},{"name":"Xixuan Hao","position":36},{"name":"Xuzhou Zhu","position":37},{"name":"Can Chen","position":38},{"name":"Jiawei Liu","position":39},{"name":"Chen Jiang","position":40},{"name":"Tao Jiang","position":41},{"name":"Fangwei Xiang","position":42},{"name":"Zepeng Yao","position":43},{"name":"Kai Chen","position":44},{"name":"Guang Yang","position":45},{"name":"Mingfei Lu","position":46},{"name":"Chunhui Li","position":47},{"name":"Zhengxuan Jiang","position":48},{"name":"Bo Tang","position":49},{"name":"Shiyu Zhan","position":50},{"name":"Yihang Sun","position":51},{"name":"Jie Chu","position":52},{"name":"Wenjun Xiong","position":53},{"name":"Zhenxiang Hua","position":54},{"name":"Chen Lin","position":55},{"name":"Detong Su","position":56},{"name":"Dongyu Deng","position":57},{"name":"Qiuping Ding","position":58},{"name":"Guofeng Ding","position":59},{"name":"Si Chen","position":60},{"name":"Guohua Wen","position":61},{"name":"Xu Chen","position":62},{"name":"Zhi-Qin John Xu","position":63},{"name":"Ying-Cong Chen","position":64},{"name":"Ningyu Zhang","position":65},{"name":"Haofen Wang","position":66},{"name":"Muning Wen","position":67},{"name":"Haotian Chen","position":68},{"name":"Wentao Zhang","position":69},{"name":"Xuanhe Zhou","position":70},{"name":"Rita Zhang","position":71},{"name":"Junchi Yan","position":72},{"name":"Feiyu Xiong","position":73}],"categories":["cs.CL","cs.LG"],"primaryCategory":"cs.CL","publishedAt":"2026-10-19T17:54:35Z","updatedAt":"2026-10-20T14:30:38Z","version":2,"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26825"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26825"},{"kind":"code","url":"https://github.com/MemTensor/MemT-1.0","label":"MemTensor/MemT-1.0"},{"kind":"project","url":"https://huggingface.co/collections/MemTensor/memt-1"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":125,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":301,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"MemTensor/MemT-1.0"},"brief":{"text":"MemT-1.0 builds memory into foundation-model computation, outperforming parametric memory baselines on average while preserving general capabilities. At 128K tokens, Nano is about 1.86 times faster and uses 66 times less persistent storage than full-context KV cache.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":3,"reason":"trending","score":0.8089},{"paper":{"id":"arxiv:2610.28354","arxivId":"2610.28354","doi":"10.48550/arXiv.2610.28354","title":"S-Agent 2.0: Spatial Reasoning with a General Coding Agent","abstract":"Spatial reasoning asks a model to work out where objects are in 3D space, how they relate to one another, and how they move. Agentic VLMs have recently brought large improvements on such tasks by letting the model actively gather evidence with 3D perception tools instead of answering in a single pass. Even so, on many questions the tools return the right evidence and the model still does not arrive at the correct answer: the gap between model and tools, where raw readings become usable conclusions, is left unfilled. We present S-Agent 2.0, a spatial reasoning agent that equips an off-the-shelf general coding agent with a command-line perception toolkit, and we design the interface around one principle: help the VLM close the evidence loop: acquire scene evidence, remember it in a persistent scene memory, and interpret each reading in plain spatial language. This design keeps the full flexibility of a coding agent, and every reading it acts on stays grounded in the tools' geometry. On MMSI-Bench and ReVSI, the toolkit improves a strong open-source model, Qwen3.8-27B, by +13.9 and +20.5 points over the same agent without tools, and a strong closed-source model, Gemini-3.8-Flash, by +11.2 and +9.6 points. With the open backbone S-Agent 2.0 reaches 66.5 on MMSI-Bench and 82.3 on ReVSI, ahead of the specialized system SpatialClaw on the same backbone (62.6 and 74.5), and with Gemini-3.8-Flash it reaches 74.1 and 90.5 against its 67.0 and 79.7. Both tool gains shrink as the backbone grows more capable, but SpatialClaw's shrinks far more on ReVSI, from +13.2 to +2.7, while ours stays at +9.6. S-Agent 2.0 also works across different general coding agent harnesses, such as Codex and Claude Code.","authors":[{"name":"Rui Gong","position":0},{"name":"Ruoyu Wang","position":1},{"name":"Yalun Dai","position":2},{"name":"Hao Li","position":3},{"name":"Sidi Yang","position":4},{"name":"Qihao Zhao","position":5},{"name":"Hongyu Li","position":6},{"name":"Zeyang Guan","position":7},{"name":"Fangzhou Hong","position":8},{"name":"Zhaoxi Chen","position":9},{"name":"Kim-Hui Yap","position":10},{"name":"Ziwei Liu","position":11}],"categories":["cs.AI","cs.CV"],"primaryCategory":"cs.AI","publishedAt":"2026-10-20T15:20:00Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.28354"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.28354"},{"kind":"code","url":"https://github.com/Ropedia/S-Agent-2.0","label":"Ropedia/S-Agent-2.0"},{"kind":"project","url":"https://ropedia.github.io/S-Agent-2.0"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.28354.jpg","label":"thumbnail"}],"signals":{"hfUpvotes":{"value":97,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":20,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"Ropedia/S-Agent-2.0"},"brief":{"text":"S-Agent 2.0 equips general coding agents with 3D perception tools and persistent scene memory, helping them turn measurements into spatial conclusions. It improves benchmark scores over tool-free agents and outperforms a specialized system across the tested backbones.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":4,"reason":"trending","score":0.7782},{"paper":{"id":"arxiv:2610.26855","arxivId":"2610.26855","doi":"10.48550/arXiv.2610.26855","title":"Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience","abstract":"World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, $Ψ_0$, $ω$-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model's utility for embodied decision-making.","authors":[{"name":"Tianqi Liu","position":0},{"name":"Hanfei Zhao","authorId":"oa:A5159042277","position":1},{"name":"Qifeng Liang","authorId":"oa:A5103721656","position":2},{"name":"Hui En Pang","authorId":"oa:A5109494610","position":3},{"name":"Jiarui Xu","position":4},{"name":"Sayang Mu","authorId":"oa:A5108969247","position":5},{"name":"Penghao Qiu","position":6},{"name":"YingSheng Zhang","position":7},{"name":"Jun Wang","position":8},{"name":"Ziang Cao","position":9},{"name":"Mingze Sun","position":10},{"name":"Yukang Cao","position":11},{"name":"Fangzhou Hong","authorId":"oa:A5031455013","position":12},{"name":"Zhaoxi Chen","position":13},{"name":"Ziwei Liu","position":14}],"categories":["cs.CV","cs.RO"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T17:59:55Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26855"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26855"},{"kind":"code","url":"https://github.com/Ropedia/ropedia-w0","label":"Ropedia/ropedia-w0"},{"kind":"project","url":"https://ropedia.github.io/ropedia-w0/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":184,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":41,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"Ropedia/ropedia-w0"},"brief":{"text":"Ropedia W0 combines large-scale human and robot egocentric data to model locomotion and manipulation across embodiments. It achieves state-of-the-art results on five benchmarks, and more human pretraining data improves performance on both human and robot tasks.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":5,"reason":"trending","score":0.7742},{"paper":{"id":"arxiv:2610.24087","arxivId":"2610.24087","doi":"10.48550/arXiv.2610.24087","title":"OR-License: Evaluating LLM Agents as Real-World Operations Research Practitioners","abstract":"Can large language model (LLM) agents deliver expert-level operations research (OR) solutions directly from original client materials, without prior expert reformulation? Existing OR benchmarks typically begin with self-contained descriptions or predefined optimization tasks, underrepresenting the translation of original business materials into professionally validated solutions. We introduce OR-License, an end-to-end benchmark of 31 instances across 10 industrial scenarios from paid OR engagements totaling USD 3.68 million. It preserves original requirements, input data, deliverable formats, and information boundaries after de-identification. To study information acquisition, we introduce ORSim, a controlled client simulator that discloses business facts without supplying OR expertise. Hidden evaluation checks feasibility and compares objective quality against original expert solutions accepted by clients before benchmark construction. Deployment-Adjusted Quality (DAQ) averages expert-relative quality across instances, assigning zero to infeasible deliveries. Across 11 frontier LLMs using coding-agent harnesses, several exceed the expert reference in conditional objective quality, but none matches its DAQ; the strongest primary Claude Code configuration achieves 88.39% strict feasibility and a DAQ of 0.87 against the expert reference of 1. Client information acquisition extends beyond clarification into solution development and validation, with some later stages showing higher confirmed issue-resolution proportions. Trajectory analysis identifies formulation and implementation failures, while task-equalized research-chain completion rates range from 24% to 45% across models on evaluable tasks. These findings motivate sustained information-gap resolution, faithful problem formulation, and evidence-linked self-improvement for reliable long-horizon agents.","authors":[{"name":"Peisong Wang","position":0},{"name":"Feixue Liu","authorId":"oa:A5147401707","position":1},{"name":"Yize Meng","authorId":"oa:A5023838790","position":2},{"name":"Bowen Liu","position":3},{"name":"Ruoyu Chen","position":4},{"name":"Gaoyang Dai","position":5},{"name":"Xing He","authorId":"oa:A5079221585","position":6},{"name":"Aiqian Li","position":7},{"name":"Fengchi Yuan","authorId":"oa:A5113310986","position":8},{"name":"Aochuan Chen","position":9},{"name":"Chenyi Zi","authorId":"oa:A5093887439","position":10},{"name":"Yuhan Li","position":11},{"name":"Hongchuan Zeng","authorId":"oa:A5102646387","position":12},{"name":"Yimo Yan","authorId":"oa:A5072530346","position":13},{"name":"Nuo Chen","position":14},{"name":"Jie Fu","position":15},{"name":"Zuo-Jun Max Shen","position":16},{"name":"Yong-Hong KUO","authorId":"oa:A5101644003","position":17},{"name":"Jia Li","position":18}],"categories":["cs.CL","cs.AI"],"primaryCategory":"cs.CL","publishedAt":"2026-10-16T18:13:25Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.24087"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.24087"},{"kind":"code","url":"https://github.com/NineAbyss/OR-License","label":"NineAbyss/OR-License"},{"kind":"project","url":"https://www.or-license.com/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":300,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":204,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"NineAbyss/OR-License"},"brief":{"text":"OR-License benchmarks LLM agents on real client materials from paid operations research engagements, finding that some beat expert solutions on objective quality when feasible, but none matches experts overall once infeasible results count as failures.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":6,"reason":"trending","score":0.7646},{"paper":{"id":"arxiv:2610.28578","arxivId":"2610.28578","doi":"10.48550/arXiv.2610.28578","title":"Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition","abstract":"In this report, we introduce Sage-A1.5, a 35B Mixture-of-Experts agentic model that builds on A1 to advance long-horizon problem solving through adaptive capability composition. Compared with A1, Sage-A1.5 learns not only to solve individual tasks, but also to organize work and revise its approach during execution. Sage-A1.5 highlights three improvements: agentic orchestration, verification, and open scientific discovery. First, agentic orchestration extends single-agent execution through task delegation and parallel collaboration among subagents with isolated contexts, enabling complex tasks to be decomposed and addressed concurrently and supporting efficient execution while reducing interference across concurrent subagents. Second, agentic verification strengthens long-horizon reliability through progress checking, evidence validation, and reflective strategy revision, allowing the agent to recognize unproductive search paths and pursue alternative approaches. Third, open scientific discovery extends agentic capabilities beyond answering predefined questions to conducting exploratory research, connecting scientific inquiry with interactive analysis and inspectable research artifacts. We showcase these capabilities through complicated agentic tasks and multidisciplinary scientific applications. Besides, we hope Sage-A1.5 provides the community with an accessible foundation for open scientific discovery, offering a practical path from long-horizon task execution to collaborative, verifiable, and inspectable scientific exploration.","authors":[{"name":"Wenjie Lou","position":0},{"name":"Tianshuo Peng","position":1},{"name":"Shiyang Feng","position":2},{"name":"Shengji Tang","position":3},{"name":"Runmin Ma","position":4},{"name":"Ruizhe Chen","position":5},{"name":"Jinxin Shi","position":6},{"name":"Haoyang Peng","position":7},{"name":"Fangchen Yu","position":8},{"name":"Zhijie Zhong","position":9},{"name":"Yuhao Zhou","position":10},{"name":"Yue Fan","position":11},{"name":"Yixin Chen","position":12},{"name":"Yiming Wu","position":13},{"name":"Yiheng Wang","position":14},{"name":"Yigeng Jiang","position":15},{"name":"Yichen Jiang","position":16},{"name":"Yi Xie","position":17},{"name":"Xu Guo","position":18},{"name":"Xiaohan He","position":19},{"name":"Tao Hu","position":20},{"name":"Songtao Huang","position":21},{"name":"Siyu Zhou","position":22},{"name":"Shuaiyu Zhang","position":23},{"name":"Shangheng Du","position":24},{"name":"Qinglong Cao","position":25},{"name":"Kun Li","position":26},{"name":"Jingqi Ye","position":27},{"name":"Hongbin Zhou","position":28},{"name":"Guohang Yan","position":29},{"name":"Zongsheng Cao","position":30},{"name":"Ziyi Wang","position":31},{"name":"Zhuo Liu","position":32},{"name":"Zhiyin Yu","position":33},{"name":"Zhiwang Zhou","position":34},{"name":"Zelin Tan","position":35},{"name":"Yuqing Zhu","position":36},{"name":"Yaoze Zhang","position":37},{"name":"Yang Chen","position":38},{"name":"Xuming He","position":39},{"name":"Xiuliang Liang","position":40},{"name":"Xingjian Guo","position":41},{"name":"Xiang Zhuang","position":42},{"name":"Weiping Shen","position":43},{"name":"Weida Wang","position":44},{"name":"Tianyuan Zhou","position":45},{"name":"Tianlin Ye","position":46},{"name":"Tengyuan Fan","position":47},{"name":"Siyu Yan","position":48},{"name":"Shuo Li","position":49},{"name":"Shuchen Pu","position":50},{"name":"Shengyuan Xu","position":51},{"name":"Shao Zhang","position":52},{"name":"Ran Wei","position":53},{"name":"Qinyuan Qin","position":54},{"name":"Qingyu Yang","position":55},{"name":"Nianchen Deng","position":56},{"name":"Meng Dai","position":57},{"name":"Kuikun Liu","position":58},{"name":"Kaitao Chen","position":59},{"name":"Juncheng Xiao","position":60},{"name":"Jiangning Liu","position":61},{"name":"Jianbiao Mei","position":62},{"name":"Jia Yu","position":63},{"name":"Haonan He","position":64},{"name":"Hangfan Zhang","position":65},{"name":"Cheng Yang","position":66},{"name":"Chen Zhang","position":67},{"name":"Bihao Zhan","position":68},{"name":"Benteng Cheng","position":69},{"name":"Yuqiang Li","position":70},{"name":"Yihao Liu","position":71},{"name":"Yang Yang","position":72},{"name":"Xujun Fu","position":73},{"name":"Xiaosong Wang","position":74},{"name":"Xiangyu Yue","position":75},{"name":"Wenwei Zhang","position":76},{"name":"Shuyue Hu","position":77},{"name":"Pinlong Cai","position":78},{"name":"Mianxin Liu","position":79},{"name":"Lei Bai","position":80},{"name":"Jie Zhou","position":81},{"name":"Jie Hou","position":82},{"name":"Jiamin Wu","position":83},{"name":"Hui Wang","position":84},{"name":"Haijun Li","position":85},{"name":"Guangyu Wang","position":86},{"name":"Dongzhan Zhou","position":87},{"name":"Chunfeng Song","position":88},{"name":"Chao Yang","position":89},{"name":"Biqing Qi","position":90},{"name":"Xing Xie","position":91},{"name":"Dahua Lin","position":92},{"name":"Bowen Zhou","position":93},{"name":"Xiangchao Yan","position":94},{"name":"Wenlong Zhang","position":95},{"name":"Shufei Zhang","position":96},{"name":"Peng Ye","position":97},{"name":"Bo Zhang","position":98}],"categories":["cs.AI"],"primaryCategory":"cs.AI","publishedAt":"2026-10-20T17:32:48Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.28578"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.28578"},{"kind":"code","url":"https://github.com/InternScience/Agents-A1","label":"InternScience/Agents-A1"},{"kind":"project","url":"https://alpha-innovator.github.io/Sage-A1.5/"}],"signals":{"hfUpvotes":{"value":65,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":574,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"InternScience/Agents-A1"},"brief":{"text":"Sage-A1.5 combines task delegation, parallel subagents, and progress checking to handle long-horizon work, revising strategies when searches stall. Its scientific applications show how agents can move beyond predefined questions to exploratory research with inspectable artifacts.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":7,"reason":"trending","score":0.7603},{"paper":{"id":"arxiv:2610.23374","arxivId":"2610.23374","doi":"10.48550/arXiv.2610.23374","title":"MOSS-Image Technical Report","abstract":"Recent closed-source image generation systems can produce diverse visual content from simple compositions to information-dense graphics, yet open-weight text-to-image models still struggle with dense text and intricate layouts. To narrow this gap, we introduce MOSS-Image, an open generation system that combines high-fidelity visual synthesis with intent-aware composition. First, we pretrain a 5B single-stream diffusion transformer (DiT) from scratch with a carefully designed recipe, enabling bilingual long-form text rendering and precise layout control. Second, we develop an intention-aware agentic prompt enhancement (PE) module that translates diverse user requests into detailed content and layout specifications. Together, they enable MOSS-Image to generate text-rich and layout-coherent images. We further introduce LongLongTextBench (LLTB) for evaluating long-form English and Chinese text rendering, and RealLayoutArena (RLA) for end-to-end layout-rich generation. MOSS-Image achieves state-of-the-art performance among open-weight models on LLTB, with an average text coverage of 96.3%. On RLA, MOSS-Image ranks second in the end-to-end system comparison, while its backbone ranks first among generators evaluated with shared enhanced prompts. We will release the model, training recipe, and evaluation framework to support further research and promote the text-to-image community.","authors":[{"name":"MOSI. AI","authorId":"oa:A5159077141","position":0}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-16T12:27:37Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23374"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23374"},{"kind":"code","url":"https://github.com/OpenMOSS/MOSS-Image","label":"OpenMOSS/MOSS-Image"},{"kind":"project","url":"https://openmoss.ai/MOSS-Image/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:04:20Z"},"signals":{"hfUpvotes":{"value":91,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":1311,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"OpenMOSS/MOSS-Image"}},"rank":8,"reason":"trending","score":0.7538},{"paper":{"id":"arxiv:2610.26828","arxivId":"2610.26828","doi":"10.48550/arXiv.2610.26828","title":"Agent World Model in the Wild: Computer-Use Agent Environment Synthesis","abstract":"Agent learning is shifting from learning from human data toward learning through interaction with the world. However, scaling such learning directly in real-world computer systems is challenging and can introduce substantial security risks. We therefore propose Agent World Model in the Wild (WildAWM), a scalable pipeline for synthesizing controllable environments and tasks for training general computer-use agents (CUAs). Whereas existing approaches typically specialize in a single agent interaction interface, we construct executable environments that jointly support graphical user interfaces, model context protocol services, and command-line interfaces, allowing agents to flexibly combine visual interaction, tool use, and code execution. WildAWM is entirely powered by open-source models, demonstrating that high-quality computer-use environments can be generated without proprietary frontier models. Beyond environments, we address three recurring challenges in task synthesis: ensuring solvability, maintaining sufficient difficulty, and producing reliable verifiers, by introducing execution-based quality gates that combine heterogeneous agent rollouts, cross-judge consistency checks, and counterfactual probes of judgment robustness. The resulting pipeline synthesizes 341 long-horizon environment-task pairs that can require over 100 agent steps to solve and remain challenging even for frontier models, with GPT-6-Astra solving only 60% of them. We train two Qwen models of different sizes exclusively in these synthetic worlds. Both models achieve substantial gains across 8 out-of-distribution benchmarks spanning web, claw, and computer-use agents, demonstrating that learning in WildAWM's synthesized worlds generalizes broadly to unseen computer-use tasks. The code is publicly available at https://github.com/MSR-Orchard/WildAWM.","authors":[{"name":"Zhaoyang Wang","position":0},{"name":"Qianhui Wu","position":1},{"name":"Haorui Wang","position":2},{"name":"Simon Yu","authorId":"oa:A5135921084","position":3},{"name":"Baolin Peng","authorId":"oa:A5066404470","position":4},{"name":"Canwen Xu","position":5},{"name":"Hao Cheng","position":6},{"name":"Wenlin Yao","authorId":"oa:A5024873469","position":7},{"name":"Boyi Liu","position":8},{"name":"Yite Wang","authorId":"oa:A5077157931","position":9},{"name":"Jianfeng Gao","position":10},{"name":"Huaxiu Yao","authorId":"oa:A5051534896","position":11}],"categories":["cs.AI","cs.CL","cs.LG"],"primaryCategory":"cs.AI","publishedAt":"2026-10-19T17:55:35Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26828"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26828"},{"kind":"code","url":"https://github.com/MSR-Orchard/WildAWM","label":"MSR-Orchard/WildAWM"},{"kind":"project","url":"https://msr-orchard.github.io/WildAWM/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":56,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":50,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"MSR-Orchard/WildAWM"},"brief":{"text":"WildAWM synthesizes controllable computer-use environments and long-horizon tasks across graphical interfaces, tools, and command lines using open-source models. Training in these worlds improved two Qwen models across eight out-of-distribution benchmarks, suggesting the synthetic tasks support broad generalization.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":9,"reason":"trending","score":0.7489},{"paper":{"id":"arxiv:2610.26625","arxivId":"2610.26625","doi":"10.48550/arXiv.2610.26625","title":"EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience","abstract":"Reconstructing human activity from egocentric video requires more than recovering individual 3D components: the observer's motion, surrounding geometry, and hand-object interactions must fit together in a consistent world. Achieving this coherence requires both joint reconstruction and aligned supervision, yet existing methods and datasets offer limited support for either. We present EgoXperience, a full-stack paradigm that addresses these needs through data collection, automatic annotation, unified modeling, and benchmarking. Central to EgoXperience is an automatic annotation engine, EgoX-Engine, that couples geometrically verified object tracking with SLAM-based world alignment, yielding consistent supervision for the camera, scene, hands, and objects. This yields EgoX-Data-6M, a 50-hour dataset of task-driven manipulation and locomotion within and across indoor and outdoor environments, with second-level task captions. We carefully select one hour across these activity categories as EgoX-Bench to evaluate joint reconstruction. To translate this supervision into reconstruction, we introduce EgoX, trained with EgoX-Data-6M and complementary annotations from additional public datasets. Given monocular egocentric video and first-frame object masks, our model achieves joint metric reconstruction of the camera, scene, hands, and objects in a shared world coordinate frame with a single forward pass. Extensive experiments demonstrate state-of-the-art performance across scene estimation and world-space hand-object pose recovery. Against the strongest evaluated baselines, EgoX reduces camera, hand, and depth errors by 68.9%, 65.6%, and 42.9%, and improves object pose scores by 28.0%.","authors":[{"name":"Xinhao Ji","authorId":"oa:A5036569224","position":0},{"name":"Hui En Pang","authorId":"oa:A5109494610","position":1},{"name":"Hao Li","position":2},{"name":"Kairui Hu","authorId":"oa:A5109778666","position":3},{"name":"Jiner Wang","authorId":"oa:A5076244968","position":4},{"name":"Mingze Sun","authorId":"oa:A5000080539","position":5},{"name":"Ziang Cao","position":6},{"name":"Yalun Dai","authorId":"oa:A5030897484","position":7},{"name":"Thang Luu","authorId":"oa:A5145226046","position":8},{"name":"Qin Ma","authorId":"oa:A5100938047","position":9},{"name":"Hongyi Chen","position":10},{"name":"Yukang Cao","position":11},{"name":"Fangzhou Hong","authorId":"oa:A5031455013","position":12},{"name":"Zhaoxi Chen","position":13},{"name":"Ziwei Liu","position":14}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T15:56:05Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26625"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26625"},{"kind":"code","url":"https://github.com/Ropedia/EgoXperience","label":"Ropedia/EgoXperience"},{"kind":"project","url":"https://ropedia.github.io/EgoXperience"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":173,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":20,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"Ropedia/EgoXperience"},"brief":{"text":"EgoXperience combines a large egocentric video dataset, automatic annotations, and a model to jointly reconstruct cameras, scenes, hands, and objects in a shared world frame. EgoX outperforms evaluated baselines, reducing camera, hand, and depth errors and improving object pose scores.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":10,"reason":"trending","score":0.7286},{"paper":{"id":"arxiv:2610.25078","arxivId":"2610.25078","doi":"10.48550/arXiv.2610.25078","title":"CoEvo-RSI: When Harnesses Evolve, Which Models Should Follow?","abstract":"Model--harness co-evolution has emerged as a prominent paradigm for recursive self-improving (RSI) agents, jointly evolving their execution scaffolds and intrinsic capabilities. We study the persistent yet often overlooked brain--body mismatch problem in this paradigm, where an evolving harness (body) demands behaviors its model (brain) has not learned or restricts capabilities it already possesses. We introduce CoEvo-RSI, a RSI framework that advances model--harness co-evolution through agent-driven, bidirectional historical inheritance. Drawing on a shared evolutionary archive and execution feedback, a meta-agent coordinates harness development from historical designs with model checkpoint selection or merging and subsequent adaptation under the selected harness, aligning inherited capabilities with evolving execution requirements to reduce brain--body mismatch. Using only supervised fine-tuning (SFT) on its own rejection-sampled trajectories, CoEvo-RSI makes Qwen3.5-9B competitive with larger models across diverse vertical applications, including multi-agent failure attribution, role-aware text-to-SQL, and smart-home instruction following. It surpasses DeepSeek-V4-Pro on TraceElephant (+14.16 agent accuracy) and GLM-5.3 on HomeBench VM (+4.90 success rate), underscoring the need for brain--body alignment and the potential of bidirectional historical inheritance to advance agent co-evolution.","authors":[{"name":"Jiaqi Su","authorId":"oa:A5148299302","position":0},{"name":"Guibin Zhang","position":1},{"name":"Jie-Jing Shao","authorId":"oa:A5087294333","position":2},{"name":"Cong Pang","authorId":"oa:A5125683346","position":3},{"name":"Tiankuo Yao","authorId":"oa:A5003449134","position":4},{"name":"Ivor Tsang","authorId":"oa:A5021751767","position":5},{"name":"Haiyan Yin","position":6},{"name":"Shuicheng Yan","position":7}],"categories":["cs.LG","cs.CL"],"primaryCategory":"cs.LG","publishedAt":"2026-10-18T07:38:03Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.25078"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.25078"},{"kind":"code","url":"https://github.com/SuJQ0722/CoEvo-RSI","label":"SuJQ0722/CoEvo-RSI"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.25078.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":127,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":5,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"SuJQ0722/CoEvo-RSI"},"brief":{"text":"CoEvo-RSI coordinates harness evolution with model selection and adaptation using shared historical designs and execution feedback, reducing brain–body mismatch. With supervised fine-tuning on its own sampled trajectories, Qwen3.5-9B outperforms larger models on two reported benchmarks.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":11,"reason":"trending","score":0.7244},{"paper":{"id":"arxiv:2610.25690","arxivId":"2610.25690","doi":"10.48550/arXiv.2610.25690","title":"ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local Navigation","abstract":"We study language-conditioned local navigation, where a robot must navigate to a text-specified target visible in its initial observation and stop at the described location. End-to-end navigation policies require diverse training data to generalize across scenes, yet real-robot data is costly to collect and simulated data leaves a sim-to-real gap. Internet egocentric videos offer a scalable source of real-world navigation experience. Our key insight is to learn goal-directed visual dynamics from Internet egocentric videos, and use these dynamics to support robot action generation. To this end, we propose ScaleNav, a unified world-action framework with three-stage training. We first perform instruction-conditioned video pretraining on these videos to learn how observations evolve as the camera approaches instructed targets across diverse real-world scenes. Action alignment then uses action-labeled simulated data to train the model to predict executable action chunks. Joint fine-tuning subsequently optimizes all trainable components together for navigation. ScaleNav achieves success rates of 86.5% on a single-view adaptation of Short-Horizon OVON and 94.9% on our HabitatGS-LN benchmark, outperforming the evaluated baselines on both benchmarks. We further deploy and evaluate ScaleNav on a real quadruped robot, with results supporting the model's ability to generalize to real-world navigation. Code and videos are available at https://zju3dv.github.io/ScaleNav/","authors":[{"name":"Yihui Yang","position":0},{"name":"Yuanhong Yu","position":1},{"name":"Jiawei Wang","position":2},{"name":"Zhongrui Yu","authorId":"oa:A5152480953","position":3},{"name":"Jia Shi","position":4},{"name":"Kun Zhan","authorId":"oa:A5129746290","position":5},{"name":"Ruizhen Hu","authorId":"oa:A5013892799","position":6},{"name":"Xiaowei Zhou","position":7},{"name":"Yu Liu","position":8},{"name":"Sida Peng","position":9}],"categories":["cs.RO"],"primaryCategory":"cs.RO","publishedAt":"2026-10-19T02:37:24Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.25690"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.25690"},{"kind":"code","url":"https://github.com/MachEmbodied/ScaleNav","label":"MachEmbodied/ScaleNav"},{"kind":"project","url":"https://zju3dv.github.io/ScaleNav/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":85,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":40,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"MachEmbodied/ScaleNav"},"brief":{"text":"ScaleNav combines instruction-guided learning from internet videos with simulated action training to generate robot navigation actions. It outperforms evaluated baselines on two benchmarks and shows real-world generalization in quadruped robot tests.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":12,"reason":"trending","score":0.723},{"paper":{"id":"arxiv:2610.26549","arxivId":"2610.26549","doi":"10.48550/arXiv.2610.26549","title":"Holi-World: Building Holistic World Intelligence from Human-Ego Experience","abstract":"Human egocentric experience offers a rich foundation for learning how the world is organized and how it changes through interaction. Realizing this potential at scale requires turning growing volumes of recordings into learning resources that connect scene structure, object semantics, and interaction histories across viewpoints and time, while reducing dependence on costly manual annotation. We introduce HOLI-WORLD, an integrated ecosystem connecting experience capture, world representation, automated data construction, model learning, and independent evaluation. To sup- port learning at scale, HOLIWORLD-Foundry combines complementary static and dynamic perception to automatically construct world records that connect persistent scene structure and evolving object states with interaction histories and supporting observations. HOLIWORLD-Data organizes these records from approximately 100 self-collected scenes into grounded learning resources, including approximately 80K QA pairs used to train HOLIWORLD-VL-4B. Independent human annotations of heldout recordings form HOLIWORLD-Bench, supporting evaluation of both annotation quality and learned spatial and temporal understanding. Experiments on ScanNet, ScanNet++, Aria Digital Twin, and HOLIWORLD-Bench demonstrate strong performance in static perception and dynamic tracking, while training on HOLIWORLD-Data substantially improves HOLIWORLD-VL-4B's spatial and temporal reasoning on HOLI- WORLD-Bench and EgoDynamic4D. HOLIWORLD provides a practical foundation for expanding learning resources from the continuous flow of human experience and advancing machine understanding of the dynamic world.","authors":[{"name":"Hao Li","position":0},{"name":"Xiyuan Zheng","position":1},{"name":"Yuanyuan Gao","position":2},{"name":"Yalun Dai","authorId":"oa:A5030897484","position":3},{"name":"Xinhao Ji","position":4},{"name":"Xuan-Li Chen","position":5},{"name":"Jing Zhao","position":6},{"name":"Fangzhou Hong","authorId":"oa:A5031455013","position":7},{"name":"Zhaoxi Chen","position":8},{"name":"Ziwei Liu","position":9}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T15:10:31Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26549"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26549"},{"kind":"code","url":"https://github.com/Ropedia/holi-world","label":"Ropedia/holi-world"},{"kind":"project","url":"https://ropedia.github.io/holi-world"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.26549.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":161,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":13,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"Ropedia/holi-world"},"brief":{"text":"HOLI-WORLD turns egocentric recordings into structured training data linking scenes, objects, and interactions, then uses it to train and evaluate a vision-language model. Training substantially improves the model’s spatial and temporal reasoning on two benchmarks.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":13,"reason":"trending","score":0.7215},{"paper":{"id":"arxiv:2610.26203","arxivId":"2610.26203","doi":"10.48550/arXiv.2610.26203","title":"Focused Omni-Modal GUI Agents","abstract":"Graphical User Interface (GUI) agents increasingly act over streaming omni-modal observations spanning screenshots, audio, video, interaction history, and device states. However, exposing this complete context uniformly at every step ignores that relevance is transient and sparse: the signal determining one action can become clutter at the next. Because executed actions feed back into the context, these errors compound into Sequential Distraction, driving the agent into irreversible trajectory drift. We formulate the underlying requirement as Focused Perception: a representation that is step-adaptive, action-sufficient, and non-intrusive. We realize this in FocusGUI, a lightweight adapter that dynamically prioritizes multimodal channels against a step-specific decision reference. Instead of uniformly processing all inputs, it compresses the most relevant evidence into a compact latent and subtly injects it into the frozen decoder's embeddings, guiding the model's attention without altering its native parameter space. FocusGUI trains from action supervision alone, requiring neither modality-relevance labels nor backbone fine-tuning. Across four omni-modal backbones, it raises overall Type Match by +33.2 and Exact Match by 3.0 times over the frozen backbone on average, and Type Match by +30.1 over the strongest existing enhancement method.","authors":[{"name":"Fuyao Zhang","position":0},{"name":"Wenwen He","position":1},{"name":"Jiaming Zhang","position":2},{"name":"Yurong Hao","position":3},{"name":"Xuan Wang","position":4},{"name":"Guowei Guan","authorId":"oa:A5108946417","position":5},{"name":"Che Wang","position":6},{"name":"Longtao Huang","position":7},{"name":"Wei Yang Bryan Lim","authorId":"oa:A5027969322","position":8}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T11:30:48Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26203"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26203"},{"kind":"code","url":"https://github.com/fyzhang1/FocusGUI","label":"fyzhang1/FocusGUI"},{"kind":"project","url":"https://fyzhang1.github.io/FocusGUI/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":227,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":12,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"fyzhang1/FocusGUI"},"brief":{"text":"FocusGUI helps GUI agents focus on the multimodal evidence most relevant to each action, without fine-tuning the backbone or labeling relevant inputs. Across four backbones, it improves average Type Match by 33.2 points over frozen models and by 30.1 over the strongest existing method.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":14,"reason":"trending","score":0.7202},{"paper":{"id":"arxiv:2610.23479","arxivId":"2610.23479","doi":"10.48550/arXiv.2610.23479","title":"HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents","abstract":"Agents built on vision language models (VLMs) now compete with learned policies on robot manipulation leaderboards. We ask how far a coding agent can go when it controls a robot instead of a computer, and where it stops. Our system, HexaShell, gives the agent a thin harness: one command line that turns its commands into robot motion and makes no decisions of its own. On the robot, the agent keeps doing what it does best: it reads output, writes code and runs tests. Two agents of the same model share the work. An acting agent runs each episode. After a failure, an optimizer agent reads the full record of the episode, with every command, image and message, finds the cause, and writes or fixes small programs called tools, each with its own tests. What the agents learn stays as code, and no demonstrations or training are needed. At run time, the tools use only what a real robot also has: RGB images, camera calibration and the state of the arms. They find objects by triangulation between camera views. Because these inputs are the same in simulation and on hardware, the tools move to a real robot with a different arm after only a small change for the arm. We build the tools on ten layouts of one seed per task with Codex and GPT-6-Astra and freeze them; the eight Open tasks share one set of tools. In RoboDojo's official evaluation on a hidden seed, 50 episodes for each of the 42 tasks, HexaShell reaches 46.9% success and a progress score of 53.2, the state of the art on the official RoboDojo leaderboard. The limit is contact, not reasoning: memory tasks reach 81.0%, but precision tasks only 32.0%. We open source the tools evolved so far at https://github.com/HexaFuture/HexaShell.","authors":[{"name":"Hongcheng Gao","authorId":"oa:A5159077373","position":0},{"name":"Alex Zhou","authorId":"oa:A5154136793","position":1},{"name":"Jay Zhu","authorId":"oa:A5101291033","position":2},{"name":"Zelin Zheng","authorId":"oa:A5136790221","position":3},{"name":"Shijia Ge","authorId":"oa:A5062661679","position":4},{"name":"Xiping Li","authorId":"oa:A5159060937","position":5},{"name":"Di Wu","authorId":"oa:A5159039113","position":6},{"name":"Yazhe Wang","authorId":"oa:A5153450943","position":7},{"name":"Jianshu Zeng","authorId":"oa:A5143767443","position":8},{"name":"Xuan Shangguan","authorId":"oa:A5153537412","position":9},{"name":"Yongli Wang","authorId":"oa:A5159029711","position":10},{"name":"Yexing Wan","authorId":"oa:A5039001935","position":11},{"name":"Lingyu He","authorId":"oa:A5153542461","position":12},{"name":"Zhiqi Jia","authorId":"oa:A5153450762","position":13},{"name":"Sihang Wu","authorId":"oa:A5059794967","position":14},{"name":"Xiao He","authorId":"oa:A5159044605","position":15}],"categories":["cs.RO"],"primaryCategory":"cs.RO","publishedAt":"2026-10-16T13:56:16Z","updatedAt":"2026-10-20T15:47:35Z","version":2,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23479"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23479"},{"kind":"code","url":"https://github.com/HexaFuture/HexaShell","label":"HexaFuture/HexaShell"},{"kind":"project","url":"https://hexafuture.ai/home/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.23479.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":314,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":98,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"HexaFuture/HexaShell"},"brief":{"text":"HexaShell lets vision-language agents control robots by writing and testing reusable code, without demonstrations or training. On RoboDojo’s hidden-seed evaluation, it achieved 46.9% success and the top leaderboard score, but struggled with precision tasks compared with memory tasks.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":15,"reason":"trending","score":0.7184},{"paper":{"id":"arxiv:2610.28630","arxivId":"2610.28630","doi":"10.48550/arXiv.2610.28630","title":"SolarRenderer: Few-Step Generative Reference Rendering","abstract":"Game engines and coding agents can specify the layout, motion, and camera of a scene, yet turning such coarse structure into realistic video requires a renderer that follows the structure and completes the appearance it leaves open. Existing renderers are tied to the representation of one engine, need tens of denoising steps or a costly distillation of their own, and learn from conditions that are not aligned with their target videos. We present SolarRenderer, a structure-conditioned video renderer that adapts a frozen MiniMax-H3 33B backbone with two lightweight LoRAs. Structural videos enter the native reference pathway of the backbone at the positions of the frames they describe, without a type identifier and described only in plain text, so that a single control LoRA follows arbitrary structural representations. Starting from a pretrained four-step LoRA, structure-guided distribution matching renders in four steps, 12.3 times faster than 50-step sampling, while preserving the learned control. A block-causal variant extends the same design to streaming and renders 768p video in real time. To train the model, we build a frame-aligned corpus of 136K clips from game engines, simulators, and YouTube videos. Experiments on an in-distribution test set and a generalization benchmark of 16 structural representations show that SolarRenderer achieves state-of-the-art results across both, and generalizes to 11 unseen representations, including pose skeletons, drawings, and point clouds.","authors":[{"name":"Guian Fang","position":0},{"name":"Shengju Qian","position":1},{"name":"Dongyue Lu","position":2},{"name":"Junchao Huang","position":3},{"name":"Lutao Jiang","position":4},{"name":"Xin Wang","position":5},{"name":"Yiren Song","position":6},{"name":"Li Jiang","position":7},{"name":"Mike Zheng Shou","position":8}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-20T17:59:48Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.28630"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.28630"},{"kind":"code","url":"https://github.com/Enderfga/SolarRenderer","label":"Enderfga/SolarRenderer"},{"kind":"project","url":"https://enderfga.cn/SolarRenderer"}],"signals":{"hfUpvotes":{"value":47,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":215,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"Enderfga/SolarRenderer"},"brief":{"text":"SolarRenderer turns plain-text-described scene structures into video using a frozen video model and lightweight adapters. It renders in four steps, supports real-time 768p streaming, and generalizes to 11 unseen structural representations.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":16,"reason":"trending","score":0.7086},{"paper":{"id":"arxiv:2610.30108","arxivId":"2610.30108","doi":"10.48550/arXiv.2610.30108","title":"Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use","abstract":"Multimodal agents (MMAs) sustain interaction through reasoning, memory, tools, and feedback, most visibly as computer-use agents, while robotic systems (RSs) couple sensing and actuation under physical dynamics. We define multimodal embodied agents (MMEAs) as goal-directed systems that couple multimodal task reasoning with physical action and revise decisions from the resulting feedback. This raises a central question: what changes when a multimodal agent moves from computer-use to robot-use? We introduce PAPAV, a unified capability-centric framework of five recurring functions: Perceive the current state, Anticipate action effects, Plan a feasible course, Act through an interface or body, and Verify the outcome. Defined by function rather than architecture, these capabilities let PAPAV identify what all three share and then compare MMEAs with MMAs and with RSs. Five physical constraints, from partial observability to unverifiable outcomes, leave fewer choices fixed in advance and less room to reverse errors. Across 64 benchmarks, explicit evaluation centers on Act; only 5 assess Anticipate and 3 assess Verify, leaving both largely hidden behind task success. These constraints also frame the open challenges we identify. PAPAV therefore provides a common basis for designing reliable physical agents and measuring progress across all five capabilities rather than task success alone.","authors":[{"name":"Yanzhe Chen","position":0},{"name":"Ziyi Yang","position":1},{"name":"Jifeng Zhu","position":2},{"name":"Qiming Huang","position":3},{"name":"Ruihe An","position":4},{"name":"Peiyao Xu","position":5},{"name":"Hesen Yang","position":6},{"name":"Runda Liu","position":7},{"name":"Chang Gong","position":8},{"name":"Zhijun Cao","position":9},{"name":"Zechen Bai","position":10},{"name":"Wenzheng Zeng","position":11},{"name":"Yiqi Lin","position":12},{"name":"Guoqiang Liang","position":13},{"name":"Kevin Yuchen Ma","position":14},{"name":"Kevin Qinghong Lin","position":15},{"name":"Mike Zheng Shou","position":16}],"categories":["cs.RO","cs.AI","cs.CV","cs.MA","cs.MM"],"primaryCategory":"cs.RO","publishedAt":"2026-10-21T17:52:48Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.30108"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.30108"},{"kind":"code","url":"https://github.com/showlab/Awesome-Multimodal-Embodied-Agent","label":"showlab/Awesome-Multimodal-Embodied-Agent"},{"kind":"project","url":"https://showlab.github.io/Awesome-Multimodal-Embodied-Agent/"}],"signals":{"hfUpvotes":{"value":14,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":137,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"showlab/Awesome-Multimodal-Embodied-Agent"},"brief":{"text":"PAPAV organizes embodied-agent capabilities into perceiving, anticipating, planning, acting, and verifying, clarifying how robot-use differs from computer-use. Across 64 benchmarks, evaluation focuses on action, while anticipation and verification remain largely unmeasured.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":17,"reason":"trending","score":0.6992},{"paper":{"id":"arxiv:2610.28577","arxivId":"2610.28577","doi":"10.48550/arXiv.2610.28577","title":"Cache Engineering for Agents","abstract":"Large Language Model (LLM) agents can handle increasingly complex tasks, navigating long trajectories of observations and actions. As these trajectories grow, key-value (KV) cache I/O and storage become the dominant cost. Trajectories can also exceed the LLM's maximum context length, requiring control over active cache growth. Prior work uses context engineering to reduce cache size by trimming or summarizing history. However, modifying the trajectory prefix invalidates subsequent cached states and triggers costly re-prefilling of retained context. Task accuracy can also suffer when too little context is retained to limit re-prefill costs. We propose cache engineering as a new scheme. It allows agents to directly control their active KV cache without modifying the trajectory or re-prefilling retained context. Specifically, agents can proactively use drop to exclude selected cache from future attention and repos to adjust retained keys' position encodings. We formalize these cache operations and identify history dependence as a serving challenge: identical token sequences can yield different cache states under different cache operation sequences, making prefix matching insufficient for correct cache reuse. Thus, we co-design the AKV serving engine. It provides a stateless interface compatible with existing serving APIs, while internally managing stateful cache reuse, eviction, and restoration. Experiments show that a tool-response rolling-drop cache policy improves average benchmark scores by 1.7 points and reduces average cost by 27.7%, compared with conventional threshold-triggered summarization. With AKV, it achieves up to 1.2x the output throughput and a 1.5x speedup in P95 request latency. Recursive self-improvement of cache policies further reduces cost by 7.1% over expert design. Our code is at https://github.com/thu-nics/AKV.","authors":[{"name":"Tianyu Fu","position":0},{"name":"Ruoxi Wang","position":1},{"name":"Zihan Min","position":2},{"name":"Zhongqi Ren","position":3},{"name":"Zhihao Li","position":4},{"name":"Tengxuan Liu","position":5},{"name":"Shengxu Mu","position":6},{"name":"Hanling Zhang","position":7},{"name":"Yu Wang","position":8}],"categories":["cs.CL"],"primaryCategory":"cs.CL","publishedAt":"2026-10-20T17:32:24Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.28577"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.28577"},{"kind":"code","url":"https://github.com/thu-nics/AKV","label":"thu-nics/AKV"},{"kind":"project","url":"https://fuvty.github.io/thinking_yard_project_page/projects/akv/"}],"signals":{"hfUpvotes":{"value":66,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":12,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"thu-nics/AKV"},"brief":{"text":"Cache engineering lets LLM agents drop or reposition cached context without changing their trajectories or re-prefilling retained history. Its AKV serving engine enables correct cache reuse, while a rolling-drop policy improves benchmark scores by 1.7 points and cuts average cost by 27.7% versus summarization.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":18,"reason":"trending","score":0.6952},{"paper":{"id":"arxiv:2610.29989","arxivId":"2610.29989","doi":"10.48550/arXiv.2610.29989","title":"Looking Ahead: Modeling Future Evolution in Autoregressive Video Generation","abstract":"Autoregressive video diffusion models generate videos chunk by chunk, yet producing plausible continuations does not guarantee that they capture how scenes evolve. Under teacher forcing, strong temporal redundancy may allow local visual continuity to serve as a shortcut for current-chunk denoising, reducing the pressure to model visual evolution. We introduce Future Chunk Prediction (FCP), an auxiliary objective jointly optimized with teacher-forced denoising that predicts temporally separated chunks without access to the intervening visual context. Drawing supervision directly from training videos, FCP requires no additional pretrained visual encoder and leaves autoregressive inference unchanged. Our method achieves superior physical plausibility over standard teacher-forced autoregressive training with a lower overall training compute budget, highlighting the effectiveness and efficiency of future-chunk supervision for modeling visual evolution. Intriguingly, our visualizations suggest that FCP fosters a more structured representation space, which may be valuable for building autoregressive world models.","authors":[{"name":"Ruopeng Gao","position":0},{"name":"Yukun Huang","position":1},{"name":"Jiarui Hu","position":2},{"name":"Shuai Wang","position":3},{"name":"Wenbo Hu","position":4},{"name":"Ying Shan","position":5},{"name":"Limin Wang","position":6}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-21T16:28:46Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.29989"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.29989"},{"kind":"code","url":"https://github.com/TencentARC/Looking-Ahead","label":"TencentARC/Looking-Ahead"},{"kind":"project","url":"https://hellorpg.github.io/looking-ahead/"}],"signals":{"hfUpvotes":{"value":62,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":3,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"TencentARC/Looking-Ahead"},"brief":{"text":"Future Chunk Prediction trains autoregressive video models to predict later frames without intervening visual context, improving physical plausibility over standard training while using less compute.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":19,"reason":"trending","score":0.6923},{"paper":{"id":"arxiv:2610.26824","arxivId":"2610.26824","doi":"10.48550/arXiv.2610.26824","title":"SpecMRL: Redesign Matryoshka Representation Learning within Parametric Spectrum","abstract":"Matryoshka Representation Learning (MRL) is designed to support flexible representation sizes by jointly optimizing nested prefixes. While effective, this design lacks an explicit constraint to ensure that the leading portion provides a suitable compact encoding. To address this, we propose SpecMRL, which reorganizes the feature space to enable more effective prefix readout. Specifically, we extend MRL with a parameter-derived spectral basis organized with feature importance. SpecMRL otherwise follows standard MRL, with prefix readout performed in the learned spectral space. Experiments on MTEB show that SpecMRL outperforms MRL at all evaluated reduced dimensions, with larger gains under more aggressive compression. These results suggest that the learned spectral basis better prioritizes informative directions, enabling more effective representation compression. Notably, SpecMRL extends the performance frontier of MRL, surpassing full-dimensional MRL with only half the embedding dimensions.","authors":[{"name":"Zhongxin Chen","authorId":"oa:A5137698410","position":0},{"name":"Songhao Wu","authorId":"oa:A5091734285","position":1},{"name":"Heng Cui","position":2},{"name":"Cong Li","position":3},{"name":"Xiao Zhou","position":4},{"name":"Rui Yan","authorId":"oa:A5003979715","position":5}],"categories":["cs.CL"],"primaryCategory":"cs.CL","publishedAt":"2026-10-19T17:54:20Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26824"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26824"},{"kind":"code","url":"https://github.com/CentreChen/Spec-MRL","label":"CentreChen/Spec-MRL"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":85,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":6,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"CentreChen/Spec-MRL"},"brief":{"text":"SpecMRL reorganizes Matryoshka representations into a learned spectral basis to make compact prefixes more informative. On MTEB, it outperforms MRL at every reduced dimension and surpasses full-dimensional MRL using half the dimensions.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":20,"reason":"trending","score":0.6919},{"paper":{"id":"arxiv:2610.20424","arxivId":"2610.20424","doi":"10.48550/arXiv.2610.20424","title":"BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business","abstract":"Existing long-horizon agent benchmarks often emphasize final task success while providing limited assessment of intermediate execution validity. Yet real-world business operations require agents to navigate extensive, potentially outdated or conflicting policies while satisfying critical constraints throughout execution. These requirements are particularly consequential in high-stakes domains such as payments, where intermediate violations can incur financial losses or regulatory consequences regardless of the final outcome. To bridge this evaluation gap, we introduce BizAgentBench, a multi-turn benchmark grounded in real-world procedures from a leading payments company. Its executable environment integrates 789 knowledge documents and 601 discoverable business tools, with consistent policies, states, and tool behavior. The benchmark comprises 50 tasks involving resource constraints, temporal dependencies, prior commitments, and conflicting or outdated guidance. Each task is validated through reference execution, while deterministic checks assess both goal attainment and execution validity without prescribing a single solution path. Experiments reveal substantial limitations even among frontier models: Claude Opus 5 and GPT-6 Sol achieve strict success rates of only 64.0% and 61.3%, respectively. For Qwen3.8-27B, goal attainment reaches 42.0%, but strict success drops to 32.7%, demonstrating how outcome-based evaluation can mask invalid execution. Beyond success rates, DeepSeek-V4.1-Flash averages 191.3 tool calls per run, approximately 70% more than reference executions, revealing substantial execution overhead. Overall, BizAgentBench establishes a testbed for evaluating and advancing long-horizon agents beyond task completion, toward reliable and constraint-compliant execution in complex business environments.","authors":[{"name":"Enhan Zhao","authorId":"oa:A5012695215","position":0},{"name":"Yongqi Tong","authorId":"oa:A5010726742","position":1},{"name":"Xuejun Yu","authorId":"oa:A5158813317","position":2},{"name":"Hang Wang","authorId":"oa:A5158794925","position":3},{"name":"Pan Wang","authorId":"oa:A5158837411","position":4},{"name":"Yichao Wang","authorId":"oa:A5158823851","position":5},{"name":"Di He","authorId":"oa:A5158796686","position":6},{"name":"Jianshe Li","authorId":"oa:A5147074669","position":7},{"name":"Wei Wu","authorId":"oa:A5158814018","position":8}],"categories":["cs.CE"],"primaryCategory":"cs.CE","publishedAt":"2026-10-14T16:16:51Z","updatedAt":"2026-10-18T15:37:22Z","version":2,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.20424"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.20424"},{"kind":"code","url":"https://github.com/ant-intl/BizAgentBench","label":"ant-intl/BizAgentBench"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.20424.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":491,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":30,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"ant-intl/BizAgentBench"},"brief":{"text":"BizAgentBench evaluates business agents on both task completion and valid execution across 50 realistic payment tasks. Even leading models often violate constraints, while outcome-only scores can hide invalid actions and inefficient tool use.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":21,"reason":"trending","score":0.6841},{"paper":{"id":"arxiv:2610.25043","arxivId":"2610.25043","doi":"10.48550/arXiv.2610.25043","title":"VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse","abstract":"Flow-matching vision-language-action (VLA) models generate an action chunk by iteratively denoising Gaussian noise, but typically execute only a short prefix before replanning with updated observations. Under this receding-horizon paradigm, the unexecuted remainder---the action tail---is discarded, and the next chunk is regenerated from fresh noise. We show that this discarded tail retains transferable denoising progress: action-tail initialization can match or outperform a native 10-step Gaussian baseline using fewer denoising steps. However, the optimal denoising budget varies substantially across samples, indicating that tail reuse should be adapted to the updated environment. Based on this observation, we introduce VLA-Rolling, a lightweight inference framework that reuses action tails through three stages: action-tail inheritance, per-action confidence estimation, and confidence-guided refinement through renoising and denoising budget allocation, while the underlying VLA remains frozen. Across $π_{0.5}$, SmolVLA, X-VLA, and MolmoAct2 on LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, VLA-Rolling reduces action-chunk inference cost in all 16 model--benchmark pairs, achieving up to a $1.48\\times$ speedup, while improving success rate in 14 pairs by up to $15.08$ percentage points. These results show that action tails are not merely discarded predictions, but reusable intermediate states that enable more efficient and accurate VLA inference. These results demonstrate that the action tail can support the generation of better actions with fewer denoising steps. Our project is available at https://chenbao123ww.github.io/VLA-Rolling/.","authors":[{"name":"Sicheng He","position":0},{"name":"Yi Pan","position":1},{"name":"Qiuting Yu","position":2},{"name":"Jiaming Huang","authorId":"oa:A5139988067","position":3},{"name":"Qi Lu","position":4},{"name":"Shuyan Wu","position":5},{"name":"Hongyan Feng","authorId":"oa:A5147481297","position":6},{"name":"Xuanyu Liu","position":7},{"name":"Jun Zhang","position":8},{"name":"Jie Zhang","position":9},{"name":"Wenqi Zhang","position":10},{"name":"Xuhong Zhang","position":11},{"name":"Jianwei Yin","position":12}],"categories":["cs.AI","cs.RO"],"primaryCategory":"cs.AI","publishedAt":"2026-10-18T06:01:43Z","updatedAt":"2026-10-21T03:43:29Z","version":3,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.25043"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.25043"},{"kind":"code","url":"https://github.com/ZJU-OmniAI/VLA-Rolling","label":"ZJU-OmniAI/VLA-Rolling"},{"kind":"project","url":"https://chenbao123ww.github.io/VLA-Rolling/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.25043.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":60,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":10,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"ZJU-OmniAI/VLA-Rolling"},"brief":{"text":"VLA-Rolling reuses discarded action tails and adapts denoising to current observations, reducing inference cost across all 16 tested model–benchmark pairs and improving success rates in 14, with up to 1.48× speedup.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":22,"reason":"trending","score":0.6805},{"paper":{"id":"arxiv:2610.28594","arxivId":"2610.28594","doi":"10.48550/arXiv.2610.28594","title":"Learning the 4D World through Point Trajectory Forecasting","abstract":"Reliable multi-view forecasting requires videos from different cameras to describe the same evolving 3D scene. To this end, we introduce Video-Point Model (VPM), a world model that conditions on multi-view video histories and a language instruction to jointly forecast future videos from arbitrary viewpoints and the 3D trajectories of persistent scene points. We extend a pretrained world model with camera-conditioned multi-view processing that integrates complementary observations across cameras, and add a trajectory head that predicts how physical points move through time. Each trajectory tracks a single point across all views, and supervising these trajectories alongside video generation therefore anchors appearance changes in every camera to a shared description of 3D motion. To support training and evaluation, we build VPM-Data, a simulation corpus pairing synchronized multi-view videos with persistent point trajectories, and VP-WorldBench, a forecasting benchmark spanning simulated and real environments. We further extend VPM to robot control by adding robot actions as a token modality in the generator. Action tokens share the point-supervised backbone and are denoised jointly with future videos, thereby grounding the policy in the same 3D-aware representation. On VP-WorldBench and existing 4D generation benchmarks, point supervision improves both video fidelity and 3D motion accuracy over a video-only counterpart, and in simulated and real-robot experiments, it also raises the manipulation success of the action-enabled model.","authors":[{"name":"Byungwoo Jeon","position":0},{"name":"Sunghwan Hong","position":1},{"name":"Dohyeon Kim","position":2},{"name":"Jonghoon Lee","position":3},{"name":"Jungwoo Park","position":4},{"name":"Inhee Lee","position":5},{"name":"Doosan Baek","position":6},{"name":"Jinil Kim","position":7},{"name":"Taehyun Lim","position":8},{"name":"Hanbyul Joo","position":9},{"name":"Marc Pollefeys","position":10},{"name":"Saining Xie","position":11},{"name":"Jinwoo Shin","position":12}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-20T17:45:45Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.28594"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.28594"},{"kind":"code","url":"https://github.com/rootyJeon/VPM","label":"rootyJeon/VPM"},{"kind":"project","url":"https://rootyjeon.github.io/video-point-model/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.28594.jpg","label":"thumbnail"}],"signals":{"hfUpvotes":{"value":43,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":22,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"rootyJeon/VPM"},"brief":{"text":"Video-Point Model jointly forecasts multi-view videos and persistent 3D point motion, using shared trajectories to connect views. Point supervision improves video fidelity, motion accuracy, and robot manipulation success over a video-only model.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":23,"reason":"trending","score":0.6597},{"paper":{"id":"arxiv:2610.28402","arxivId":"2610.28402","doi":"10.48550/arXiv.2610.28402","title":"StreamEval: Evaluating Interactive Streaming Video Generation Models","abstract":"As streaming video generation advances, enabling real-time interaction with generated video is becoming increasingly important. We introduce StreamEval, a unified benchmark for interactive streaming video generation spanning five tasks: avatar dialogue, avatar performance, world generation, instruction editing, and reference editing. StreamEval comprises 300 scripted test cases lasting 15-300s, each specifying initial assets, incrementally arriving multimodal requests, and explicit requirements for state changes and content preservation. Performance is assessed along four evaluation dimensions: task execution, state consistency, visual fidelity, and audiovisual coordination, instantiated through ten task-specific metrics on a 0-100 scale. Evaluating 37 distinct models across 300 test cases, we find that high visual quality and identity retention do not necessarily coincide with reliable request execution or sustained state consistency. In particular, request execution remains challenging: the highest task-specific scores are 36.90 for Following among seven avatar-performance models, 25.93 for Following in text-conditioned world generation, and 15.00 for Edit Success among three reported reference-editing models. These findings identify reliable request execution under timing and preservation constraints as a central challenge for advancing omni-modal real-time interactive systems. The leaderboard is available at https://streameval.ai","authors":[{"name":"StreamEval Team","position":0}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-20T15:47:43Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.28402"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.28402"},{"kind":"code","url":"https://github.com/StreamEval/StreamEval","label":"StreamEval/StreamEval"},{"kind":"project","url":"https://streameval.ai/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.28402.jpg","label":"thumbnail"}],"signals":{"hfUpvotes":{"value":32,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":129,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"StreamEval/StreamEval"},"brief":{"text":"StreamEval benchmarks 37 models across five interactive streaming video tasks, assessing request execution, state consistency, visual fidelity, and audiovisual coordination. Results show that strong visual quality and identity retention do not guarantee reliable request execution, which remains a central challenge.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":24,"reason":"trending","score":0.659},{"paper":{"id":"arxiv:2610.24960","arxivId":"2610.24960","doi":"10.48550/arXiv.2610.24960","title":"CoworkEnv: Synthesizing Training Environments for Long-Horizon Cowork Agents","abstract":"AI coworkers must coordinate tools, communicate across turns, and act on changing workspace state while respecting workflow constraints. Constructing training experience for these capabilities requires coherent requests, service records, tool interfaces, and grading checks. We present CoworkEnv, a capability-guided framework for synthesizing executable training environments for AI coworkers. Workflow requirements guide task construction, and programmatic validation with feedback repair checks generated specifications and grading behavior. Teacher executions yield curated demonstrations for supervised fine-tuning (SFT), while student interactions provide rollouts and rewards for reinforcement learning (RL). Using these resources, we train compact PocketWorker models. For Qwen3-VL-8B-Instruct, SFT raises the mean score from 46.6 to 64.1 on Claw-Eval, including a 31.9-point gain on Multi-turn tasks. An SFT-only extension to ClawMark demonstrates applicability across coworker benchmarks. We further compare RL training dynamics across configurations that vary reward processing and credit assignment. Improving training rewards provide preliminary evidence that the constructed environments support policy optimization.","authors":[{"name":"Zhongwei Xie","position":0},{"name":"Yiming Du","position":1},{"name":"Jiaxin Bai","authorId":"oa:A5026258275","position":2},{"name":"Baixuan Xu","position":3},{"name":"Chunyang Li","authorId":"oa:A5146889494","position":4},{"name":"Xiyu Ren","position":5},{"name":"Wei Fan","authorId":"oa:A5158750280","position":6},{"name":"Tianshi Zheng","authorId":"oa:A5031583104","position":7},{"name":"Weizhi Du","authorId":"oa:A5011575160","position":8},{"name":"Jingqiao Yu","authorId":"oa:A5104162778","position":9},{"name":"Xingsu Li","authorId":"oa:A5143956373","position":10},{"name":"Zhaowei Wang","position":11},{"name":"Yangqiu Song","authorId":"oa:A5158711285","position":12}],"categories":["cs.AI","cs.CL"],"primaryCategory":"cs.AI","publishedAt":"2026-10-18T02:20:30Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.24960"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.24960"},{"kind":"code","url":"https://github.com/HKUST-KnowComp/CoworkEnv","label":"HKUST-KnowComp/CoworkEnv"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.24960.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":47,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":12,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"HKUST-KnowComp/CoworkEnv"},"brief":{"text":"CoworkEnv synthesizes executable training environments for AI coworkers, using validated workflows to generate demonstrations and reinforcement-learning data. Training raises Qwen3-VL-8B-Instruct’s Claw-Eval mean score from 46.6 to 64.1, with a 31.9-point gain on multi-turn tasks.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":25,"reason":"trending","score":0.6574},{"paper":{"id":"arxiv:2610.17665","arxivId":"2610.17665","doi":"10.48550/arXiv.2610.17665","title":"RoTA-Bench: Evaluating 3D Spatial Reasoning through Interactive Projection Alignment","abstract":"Can vision-language models turn 2D observations into 3D understanding that supports precise action? This requires inferring object geometry, anticipating the effects of rotation, and refining actions through visual feedback. We introduce RoTA-Bench to evaluate these abilities through interactive 3D-to-2D projection alignment. Given a target silhouette, a model observes a rotated object and chooses successive rotations to recover the target projection within a limited action budget. Our automated pipeline constructs silhouette-preserving 3D objects from 2D icons and uses reversible initialization to guarantee solvability, yielding 674 tasks and a stratified 100-task test set. Empirically, we find that (i) a substantial human-model gap persists: GPT-5.6-Sol succeeds on only 17% of tasks, whereas 20 undergraduate participants collectively complete the test set four times with a 100% success rate; (ii) image-space and pose-space progress can diverge: at matched states, the model more often improves silhouette overlap while worsening pose alignment than humans (19.4% vs. 10.9%); and (iii) fine-grained control is sensitive to interaction history: removing it leaves 96.8% of the model's actions at a coarse 30-degree step size and sharply reduces observed task success. These findings highlight the challenge of converting visual evidence into precise 3D control, beyond recognizing or matching individual images.","authors":[{"name":"Qingchuan Ma","authorId":"oa:A5084762321","position":0},{"name":"Yuexiao Ma","authorId":"oa:A5158724314","position":1},{"name":"Xiang Meng","authorId":"oa:A5158691939","position":2},{"name":"Yixuan Zou","authorId":"oa:A5129903955","position":3},{"name":"Tianyu Xie","authorId":"oa:A5006971320","position":4},{"name":"Wang Chen","authorId":"oa:A5158717947","position":5},{"name":"Yuhui Zeng","authorId":"oa:A5158688253","position":6},{"name":"Yuhang Wu","authorId":"oa:A5158715993","position":7},{"name":"Rongrong Ji","authorId":"oa:A5159089594","position":8},{"name":"Xiawu Zheng","authorId":"oa:A5054226277","position":9}],"categories":["cs.AI"],"primaryCategory":"cs.AI","publishedAt":"2026-10-13T03:46:48Z","updatedAt":"2026-10-16T07:18:17Z","version":3,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.17665"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.17665"},{"kind":"code","url":"https://github.com/MAC-AutoML/RoTA-Bench","label":"MAC-AutoML/RoTA-Bench"},{"kind":"project","url":"https://www.modelscope.cn/studios/san2333333/rotabench-anon?mode=full&app_path=%2Fportal"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.17665.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":207,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":6,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"MAC-AutoML/RoTA-Bench"}},"rank":26,"reason":"trending","score":0.653},{"paper":{"id":"arxiv:2610.22791","arxivId":"2610.22791","doi":"10.48550/arXiv.2610.22791","title":"Self-Evolving Coding Agents via Harness-Policy Optimization","abstract":"Large language model (LLM) agents have shown promising potential in coding tasks, and their performance is jointly determined by the policy model and the harness that mediates their interaction with the environment. Indeed, harness and policy are mutually reinforcing and evolving only one component is suboptimal---a harness tuned for one checkpoint may not target the failure modes of an updated policy, while policy learning remains limited to the trajectories elicited by a fixed harness. In light of this, we propose Self-HPO, a recursive self-improving framework for coding agents that alternates harness and policy optimization, using the evolving policy as its own harness proposer. Specifically, in each round, the current policy first analyzes its own failures and proposes new harness candidates (H-step). The policy is then updated through reinforcement learning (RL) under the selected harness (P-step), and hard tasks identified during training form the next harness search set, keeping harness optimization targeted to the policy's evolving failure modes. Across three coding benchmarks, Self-HPO consistently outperforms harness-only and policy-only baselines, improving Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8%, with gains generalizing to out-of-distribution settings. Further analysis shows that both optimized components retain gains even when paired with their base counterparts, while their joint gain exceeds the sum of their individual gains. Self-HPO also improves reliability (Pass@1) and expands the capability frontier (Pass@k) over the base agent. Additionally, we show that with a stronger external model (e.g., Opus 4.8) as the harness proposer, the same framework yields a further performance boost.","authors":[{"name":"Zhepei Wei","authorId":"oa:A5159037868","position":0},{"name":"Wenlin Yao","authorId":"oa:A5024873469","position":1},{"name":"Baolin Peng","authorId":"oa:A5066404470","position":2},{"name":"Danqing Wang","authorId":"oa:A5127795602","position":3},{"name":"Qianhui Wu","authorId":"oa:A5159090754","position":4},{"name":"Hao Cheng","authorId":"oa:A5159032636","position":5},{"name":"Yu Meng","authorId":"oa:A5136000161","position":6},{"name":"Jianfeng Gao","authorId":"oa:A5159026991","position":7}],"categories":["cs.CL","cs.AI"],"primaryCategory":"cs.CL","publishedAt":"2026-10-16T01:07:52Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.22791"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.22791"},{"kind":"code","url":"https://github.com/MSR-Orchard/Self-HPO","label":"MSR-Orchard/Self-HPO"},{"kind":"project","url":"https://weizhepei.com/blog/self-hpo/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.22791.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":47,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":30,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"MSR-Orchard/Self-HPO"}},"rank":27,"reason":"trending","score":0.6507},{"paper":{"id":"arxiv:2610.23121","arxivId":"2610.23121","doi":"10.48550/arXiv.2610.23121","title":"Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation","abstract":"Generating complete, coherent, and usable 3D indoor scenes from natural language requires strong \\textbf{design intelligence}: the ability to design high-quality scenes and critique their quality. However, existing approaches either rely on general-purpose models lacking domain-specific design experience or learn from static scenes. In this work, we introduce \\textbf{Scene-Designer}, an agentic framework that learns design intelligence through complementary design and critique experiences. We construct two datasets: \\textbf{Scene-OTA-35K}, which provides Observation--Thought--Action supervision for learning scene design and efficient error correction, and \\textbf{Scene-OTC-15K}, which provides Observation--Thought--Critic supervision for learning scene quality critique across physical validity, functional correctness, and layout coherence. Based on these experiences, we propose \\textbf{Scene-GRPO}, a Critic-guided GRPO framework that improves design decisions and reduces unnecessary trial-and-error corrections. Extensive experiments demonstrate that Scene-Designer achieves superior performance in physical validity, functional correctness, and layout coherence, while the learned Scene-Critic significantly outperforms general-purpose models in scene critique. Moreover, test-time Critic guidance further improves scene generation, showing that the Critic can provide useful guidance beyond training. Additional analysis suggests a trend of mutual enhancement between scene design and critique abilities.","authors":[{"name":"Xuanyu Liu","authorId":"oa:A5159059397","position":0},{"name":"Hongyan Feng","authorId":"oa:A5147481297","position":1},{"name":"Miao Pan","authorId":"oa:A5122915476","position":2},{"name":"Sunlai Chen","authorId":"oa:A5147455532","position":3},{"name":"Lizhou Liang","authorId":"oa:A5091685008","position":4},{"name":"Bing Sun","authorId":"oa:A5159041702","position":5},{"name":"Wenqi Zhang","authorId":"oa:A5159035467","position":6},{"name":"Xuhong Zhang","authorId":"oa:A5122200558","position":7}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-16T08:28:14Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23121"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23121"},{"kind":"code","url":"https://github.com/ZJU-OmniAI/Scene-Designer","label":"ZJU-OmniAI/Scene-Designer"},{"kind":"project","url":"https://zju-omniai.github.io/Scene-Designer/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":63,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":27,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"ZJU-OmniAI/Scene-Designer"}},"rank":28,"reason":"trending","score":0.6468},{"paper":{"id":"arxiv:2610.22169","arxivId":"2610.22169","doi":"10.48550/arXiv.2610.22169","title":"Lexical Reference Frames: Interpreting LM Residual Streams in Native Semantics","abstract":"The vocabulary of a language model is its native semantics, and the embedding and unembedding fix the meaning of a state at the two ends of the residual stream. Between them, existing methods reuse a fixed lens or treat states as a point cloud, and neither reads a state as the model does. We argue that every hidden state has a native reference frame, induced by the layers that remain and adaptive to the context. Lexical reference frames approximate it in a model-native and context-aware way, reading a state through the remaining layers with a given number of rounds of attention between positions and a typical attention output for the rest, fitting nothing to the model's output. No lens with fixed directions per layer can reach a mean cosine with the native frame's lexical readouts above the square root of their coherence, 0.30 and 0.35 at mid-depth of SmolLM3-3B and OLMo3-7B. LexFrame[1], of order 1, reaches 0.97 and 0.97, and its readings are on average within 0.16 and 0.15 nats of the model's output. Its coefficient differences attribute predictions more faithfully than attribution patching, and its lexical basis steers predictions while sparing other tokens. On average, its readings settle on the model's top token over the final 0.82 and 0.81 of the layers, those of lenses over at most 0.11. These frames thus make every hidden state readable, attributable and steerable.","authors":[{"name":"Jian Gu","authorId":"oa:A5158927886","position":0},{"name":"Hongyu Zhang","authorId":"oa:A5158855644","position":1},{"name":"Chunyang Chen","authorId":"oa:A5158904488","position":2},{"name":"Aldeida Aleti","authorId":"oa:A5035346540","position":3}],"categories":["cs.CL","cs.LG"],"primaryCategory":"cs.CL","publishedAt":"2026-10-15T16:29:27Z","updatedAt":"2026-10-18T11:22:37Z","version":2,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.22169"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.22169"},{"kind":"code","url":"https://github.com/jianguda/lexframe","label":"jianguda/lexframe"},{"kind":"project","url":"https://jianguda.github.io/blogs/lexframe/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:04:20Z"},"signals":{"hfUpvotes":{"value":56,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":6,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"jianguda/lexframe"},"brief":{"text":"Lexical reference frames interpret hidden states using the model’s remaining layers and context, rather than a fixed lens. They closely match the model’s lexical readouts, settle on its top prediction earlier than lenses, and support more faithful attribution and targeted steering.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":29,"reason":"trending","score":0.6354},{"paper":{"id":"arxiv:2610.27999","arxivId":"2610.27999","doi":"10.48550/arXiv.2610.27999","title":"Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds","abstract":"Moving from controllable videos to genuinely playable worlds requires persistent world state, executable game rules, and low-latency interaction. We outline a roadmap for the next generation of Matrix-Game that separates game execution from visual generation through an executable game harness, and introduce Matrix-Game 4.0 as the first system built on this design. A lightweight game engine maintains world state and executes interaction rules, a multimodal coding agent translates player ideas into game code and dynamic scene captions, and a renderer based on autoregressive video model transforms engine-rendered low-poly proxies into detailed visuals, conditioned on dynamic captions and memory references. This framework provides continuous guidance for visual generation and enables natural-language creation and modification of scenes and gameplay. Our renderer is trained on aligned low-poly and high-fidelity video pairs and distilled into a causal generator that generates in four-frame chunks with one or two denoising steps. Together with targeted inference optimizations, Matrix-Game 4.0 enables 480p, 60 FPS gameplay on a single NVIDIA RTX5090 with a key-press-to-display latency of 112ms. Single-step generation achieves 83FPS steady-state throughput with 18.2GiB of peak GPU memory usage. With fixed prompts, steady-state throughput also reaches 67FPS on an RTX5080 and 58FPS on an RTX5070Ti. We demonstrate customizable shooter and driving games with engine-enforced mechanics, and release code, model weights, and low-poly game assets to support research and creation of playable game worlds.","authors":[{"name":"Jihai Zhang","position":0},{"name":"Runjia Qian","position":1},{"name":"Kaiqin Kong","position":2},{"name":"Biao Jiang","position":3},{"name":"Puyi Wang","position":4},{"name":"Siru Chen","position":5},{"name":"Tianle Li","position":6},{"name":"Guanjie Chen","position":7},{"name":"Junyi Wu","position":8},{"name":"Chubin Chen","position":9},{"name":"Yuanhang Li","position":10},{"name":"Kaichen Huang","position":11},{"name":"Haobo Zhang","position":12},{"name":"Yangguang Li","position":13},{"name":"Wei Yu","position":14},{"name":"Guibin Chen","position":15},{"name":"Yu Cheng","position":16},{"name":"Yahui Zhou","position":17}],"categories":["cs.CV","cs.GR"],"primaryCategory":"cs.CV","publishedAt":"2026-10-20T10:41:38Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.27999"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.27999"},{"kind":"project","url":"https://matrix-game-v4-0.github.io/page/"}],"signals":{"hfUpvotes":{"value":183,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"}},"brief":{"text":"Matrix-Game 4.0 combines an executable game engine, a coding agent, and a video renderer to create and modify playable worlds from natural language. It delivers 480p gameplay at 60 FPS on an RTX 5090, with engine-enforced mechanics.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":30,"reason":"trending","score":0.6278},{"paper":{"id":"arxiv:2610.23279","arxivId":"2610.23279","doi":"10.48550/arXiv.2610.23279","title":"Xiaomi-Lens-0 Technical Report","abstract":"Multimodal large language models (MLLMs) answer questions about whole images well, yet they still struggle with fine-grained region understanding. A key reason is that text-only supervision never specifies which regions support an answer. To close this gap, we present Xiaomi-Lens-0, a 4B MLLM built on Qwen3.5 that predicts masks from its own visual representations. Specifically, a lightweight decoder combines vision-encoder patch features with the language model's mask-token states, termed visual latents, so that mask supervision reaches the shared backbone without an additional segmentation encoder. Because each latent decodes to a mask, the model can interleave reasoning with region-specific latents and thus think with visual latents, while grounded reasoning traces supervise the intermediate regions referenced during chain-of-thought. As a result, across all eight RefCOCO, RefCOCO+, and RefCOCOg splits, Xiaomi-Lens-0 gains 5.6 points on average over Qwen3.5 in referring comprehension and achieves the highest segmentation cIoU among compared models. Moreover, it surpasses Qwen3.5 across all perception benchmarks. Homepage: https://seerray-lab.github.io/Xiaomi-Lens-0","authors":[{"name":"Pei Fu","authorId":"oa:A5159028408","position":0},{"name":"Yike Liu","authorId":"oa:A5159066069","position":1},{"name":"Yue Sun","authorId":"oa:A5159029279","position":2},{"name":"Hanmo Wang","authorId":"oa:A5051466085","position":3},{"name":"Xiuwen Xi","authorId":"oa:A5123963621","position":4},{"name":"Ruoceng Zhang","authorId":"oa:A5159087642","position":5},{"name":"Shaojie Zhang","authorId":"oa:A5159040864","position":6},{"name":"Tingwei Huang","authorId":"oa:A5138704077","position":7},{"name":"Chao Jiang","position":8},{"name":"Hang Li","position":9},{"name":"Bingrui Liu","authorId":"oa:A5159079181","position":10},{"name":"Zhenbo Luo","position":11},{"name":"Heng Qu","position":12},{"name":"Changjian Xu","authorId":"oa:A5159040432","position":13},{"name":"Ming Xu","position":14},{"name":"Jian Luan","authorId":"oa:A5054843960","position":15}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-16T10:50:41Z","updatedAt":"2026-10-19T11:43:25Z","version":2,"license":"http://creativecommons.org/publicdomain/zero/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23279"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23279"},{"kind":"code","url":"https://github.com/SeerRay-Lab/Xiaomi-Lens-0","label":"SeerRay-Lab/Xiaomi-Lens-0"},{"kind":"project","url":"https://seerray-lab.github.io/Xiaomi-Lens-0"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.23279.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":47,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":47,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"SeerRay-Lab/Xiaomi-Lens-0"},"brief":{"text":"Xiaomi-Lens-0 adds mask prediction and region-grounded reasoning to a 4B multimodal model, enabling it to connect answers to image regions. It improves on Qwen3.5 across all perception benchmarks and leads compared models in segmentation cIoU.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":31,"reason":"trending","score":0.6251},{"paper":{"id":"arxiv:2610.26659","arxivId":"2610.26659","doi":"10.48550/arXiv.2610.26659","title":"SpatialScout: Can VLM Agents Actively Explore and Map 3D Scenes?","abstract":"Spatial intelligence requires agents to actively acquire visual evidence and integrate observations into a coherent representation of the three-dimensional world. Yet spatial question answering offers limited evidence of whether models can construct accurate metric maps through autonomous exploration. We introduce SpatialScout, a benchmark for evaluating vision-language model (VLM) agents on active object-level 3D mapping. SpatialScout comprises 80 diverse indoor scenes containing 3,140 target objects. Given an object inventory, RGB observations, camera calibration, and camera poses, agents control a camera to explore each scene and estimate object centers in a shared coordinate frame. The task requires coordinating viewpoint selection, cross-view association, geometric estimation, and map revision under bounded interaction budgets. Evaluation jointly measures localization accuracy and target coverage through object-scale overlap scores that penalize missing predictions. Experiments with frontier VLMs reveal substantial limitations: the strongest evaluated model, Claude Opus 5.5, achieves a mean overlap score of 41.71%, with 47.23% of targets exceeding an intersection-over-union (IoU) threshold of 0.5 and only 3.25% exceeding 0.9. These results reveal a substantial gap between recovering coarse object layouts and constructing complete, precise metric maps. SpatialScout provides a quantitative testbed for developing agents that build reliable 3D spatial representations through active perception. Code is available at https://github.com/JesseZZZZZ/SpatialScout.","authors":[{"name":"Zhaoxi Zhang","authorId":"oa:A5126598152","position":0},{"name":"Hao Liang","position":1},{"name":"Linxuan Wu","authorId":"oa:A5133671231","position":2},{"name":"Yiming Wu","position":3},{"name":"Yifu Yuan","position":4},{"name":"Kun Liang","position":5},{"name":"Chenming Tang","position":6},{"name":"Meiyi Qiang","authorId":"oa:A5120132025","position":7},{"name":"Wentao Zhang","position":8},{"name":"Weikang Li","authorId":"oa:A5122392521","position":9}],"categories":["cs.CV","cs.CL"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T16:17:48Z","updatedAt":"2026-10-20T02:49:48Z","version":2,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26659"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26659"},{"kind":"code","url":"https://github.com/JesseZZZZZ/SpatialScout","label":"JesseZZZZZ/SpatialScout"},{"kind":"project","url":"https://spatialscout.osslab-pku.org"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.26659.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":38,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":2,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"JesseZZZZZ/SpatialScout"},"brief":{"text":"SpatialScout benchmarks VLM agents’ ability to explore indoor scenes and map objects in 3D; the best tested model achieves 41.71% mean overlap, revealing a gap between coarse layouts and complete, precise maps.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":32,"reason":"trending","score":0.6246},{"paper":{"id":"arxiv:2610.25404","arxivId":"2610.25404","doi":"10.48550/arXiv.2610.25404","title":"MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion","abstract":"Multi-turn image editing aims to modify an image according to a sequence of instructions, enabling more effective interactive visual content creation for users. However, existing methods often exhibit noticeable drift after only a few editing turns, manifesting in two aspects: 1) unstable visual quality, with accumulating artifacts and degraded details; and 2) inconsistent image content, with unintended changes to preserved regions and prior edits. To address these challenges, we propose MT Forcing, a novel training paradigm for multi-turn image editing with unified multimodal models (UMMs) based on autoregressive diffusion. Specifically, we propose Single-Turn Distribution Matching Distillation (Single-Turn DMD), which supervises student self-rollouts with a teacher that applies cumulative edits to the source image in a single turn, providing visual-quality guidance without recursive error accumulation. Additionally, we introduce Teacher-Forcing Guidance, which supervises history-dependent transitions on reference trajectories during warmup and as a regularizer during distillation, encouraging cross-turn content consistency beyond what per-turn distribution matching alone can ensure. Extensive experiments demonstrate that our approach significantly improves the multi-turn image editing capabilities of UMMs, yielding more stable visual quality and stronger cross-turn consistency.","authors":[{"name":"Houyuan Chen","authorId":"oa:A5159041480","position":0},{"name":"Fanzhou Wang","authorId":"oa:A5021955233","position":1},{"name":"Xiangyu Fan","position":2},{"name":"Zesong Qiu","authorId":"oa:A5046875669","position":3},{"name":"Yukun Wei","position":4},{"name":"Songlin Yang","position":5},{"name":"Zhitao Yang","authorId":"oa:A5119736214","position":6},{"name":"Lvmin Zhang","authorId":"oa:A5079385579","position":7},{"name":"Pei Yang","position":8},{"name":"Hao Zhao","position":9},{"name":"Dahua Lin","position":10},{"name":"Lei Yang","position":11},{"name":"Anyi Rao","authorId":"oa:A5067715162","position":12}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-18T17:01:40Z","updatedAt":"2026-10-21T08:45:58Z","version":2,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.25404"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.25404"},{"kind":"code","url":"https://github.com/houyuanchen111/MT-Forcing","label":"houyuanchen111/MT-Forcing"},{"kind":"project","url":"https://houyuanchen111.github.io/mt-forcing.github.io/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":41,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":22,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"houyuanchen111/MT-Forcing"},"brief":{"text":"MT Forcing trains autoregressive diffusion models to edit images across multiple instruction turns, using single-turn quality guidance and reference trajectories to preserve prior edits. Experiments show more stable visual quality and stronger consistency across turns.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":33,"reason":"trending","score":0.6197},{"paper":{"id":"arxiv:2610.28698","arxivId":"2610.28698","doi":"10.48550/arXiv.2610.28698","title":"Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds","abstract":"Scaling data, diversity and model capacity has improved embodied learning in physical environments, where diverse experiences remain grounded in shared physical laws. Whether these gains extend to virtual worlds with different visual styles, dynamics, controls, and objectives remains unclear. We investigate this question through a controlled study of continued pre-training (CPT) across three axes: data volume, diversity, and model size. Our 120 training configurations span four model sizes from 0.8B to 27B, six data budgets from 50 to 1,600 hours, and five mixtures ranging from Minecraft-only data to diverse multi-world corpora. We assess performance and generalization using three-level validation losses and gameplay benchmarks from in-corpus worlds to unseen worlds. Our results show: (1) more data improves in-corpus performance, diversity improves transfer to related worlds with limited gains in distant worlds, and model size improves data efficiency but generalizes worse to distant worlds under single-world training; (2) we observe an early interface phase with broad cross-world benefits and a later world phase with gains narrowed to in-corpus and related worlds, and fit an empirical scaling law relating validation loss to data volume, diversity and model size; and (3) our models are competitive with Minecraft specialists and frontier models in 3D worlds, and perform best in real-time settings. However, the limits of imitation and persistent gaps in distant worlds suggest that building generalist embodied agents requires richer training stages beyond continued pre-training.","authors":[{"name":"Kuan Zhang","position":0},{"name":"Yukun Chen","position":1},{"name":"Zhihao Yang","position":2},{"name":"Yue Su","position":3},{"name":"Xiangnan Wu","position":4},{"name":"Zirong Chen","position":5},{"name":"Run Luo","position":6},{"name":"Jinkun Hou","position":7},{"name":"Tao Tan","position":8},{"name":"Yinhe Zheng","position":9},{"name":"Yiming Li","position":10}],"categories":["cs.AI","cs.CV","cs.LG"],"primaryCategory":"cs.AI","publishedAt":"2026-10-20T11:09:09Z","version":1,"license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.28698"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.28698"},{"kind":"code","url":"https://github.com/THUSI-Lab/GameScaling","label":"THUSI-Lab/GameScaling"},{"kind":"project","url":"https://gamescaling.github.io"}],"signals":{"hfUpvotes":{"value":17,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":9,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"THUSI-Lab/GameScaling"},"brief":{"text":"Continued pre-training across virtual worlds shows that more data boosts in-world performance, while diverse data helps transfer mainly to related worlds. Larger models use data more efficiently, but training on a single world can hurt transfer to distant worlds.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":34,"reason":"trending","score":0.6149},{"paper":{"id":"arxiv:2610.26851","arxivId":"2610.26851","doi":"10.48550/arXiv.2610.26851","title":"Manga2Animate: An Agentic Framework for Faithful Manga Animation with Script and Storyboard Planning","abstract":"Adapting existing manga into coherent animation that stays faithful to the source is a valuable yet challenging task. Existing methods either access the manga mainly through text, losing its characters and art style, or feed manga pages directly into video generation models, which misinterpret the manga's structure and produce incorrect storylines. In this paper, we propose Manga2Animate, a multi-agent framework that grounds every stage of generation in the source manga. We decompose the task into two intermediate representations, an animation script and a visual storyboard. Each representation is produced by a creator agent and reviewed against the source panels by a paired reviewer agent. We tailor representation-specific selection and dual-level feedback to progressively refine both before synthesizing the final animation. Extensive experiments and a user study demonstrate that Manga2Animate produces animations more faithful to the source manga and more consistent across shots than state-of-the-art approaches. Project page: https://annchenn.github.io/Manga2Animate/","authors":[{"name":"Sheng-Wei Zhang","position":0},{"name":"Hsi-An Chen","authorId":"oa:A5157686593","position":1},{"name":"Yan-Ting Liou","authorId":"oa:A5103468431","position":2},{"name":"Chih-Hsuan Chen","authorId":"oa:A5157775567","position":3},{"name":"Shih-Chen Tseng","authorId":"oa:A5087520959","position":4},{"name":"Yu-Lun Liu","position":5}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T17:59:46Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26851"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26851"},{"kind":"code","url":"https://github.com/annchenn/Manga2Animate","label":"annchenn/Manga2Animate"},{"kind":"project","url":"https://annchenn.github.io/Manga2Animate/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.26851.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":29,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":51,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"annchenn/Manga2Animate"},"brief":{"text":"Manga2Animate uses paired creator and reviewer agents to refine animation scripts and storyboards against manga panels before generating animation. Experiments and a user study find its animations are more faithful to the source and more consistent across shots than leading approaches.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":35,"reason":"trending","score":0.6131},{"paper":{"id":"arxiv:2610.26844","arxivId":"2610.26844","doi":"10.48550/arXiv.2610.26844","title":"AnyGoal: On World Model Generalization Beyond Expert Rollouts","abstract":"The goal of a world model is to capture the dynamics of an environment. This has the great appeal that it allows agents to plan and simulate a course of actions with the help of a world model. For these uses to be reliable, however, the learned dynamics must behave well beyond the specific state-action combinations observed in expert rollouts. Despite this, the relationship between data coverage, dynamics generalization, and downstream planning performance remains underexplored. We study this relationship through the lens of compositional generalization: can a world model recombine familiar state and action factors into valid dynamics? To answer this, we introduce AnyGoal, a benchmark of goal-shifted task variants built from widely used world-model environments, together with an open-source library. Each task is designed to be within reach of a world model that has learned the underlying dynamics: individual state factors and actions remain within their training support, but their combinations are underrepresented in expert trajectories. Across multiple environments and world model families, we find that current models consistently exhibit failures under these recombinations. By examining dynamics and representation failures across model families, interaction regimes, and evaluation metrics, we study where current world models generalize and where challenges remain for goal-shifted prediction and planning. Together, this framework, benchmark, and analysis provide a basis for evaluating progress toward world models that generalize to new planning tasks.","authors":[{"name":"Heejeong Nam","authorId":"oa:A5054896116","position":0},{"name":"Divyat Mahajan","authorId":"oa:A5082071152","position":1},{"name":"Quentin Le Lidec","position":2},{"name":"Randall Balestriero","authorId":"oa:A5047293370","position":3},{"name":"David Klindt","authorId":"oa:A5059395392","position":4}],"categories":["cs.AI"],"primaryCategory":"cs.AI","publishedAt":"2026-10-19T17:59:05Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26844"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26844"},{"kind":"code","url":"https://github.com/galilai-group/AnyGoal","label":"galilai-group/AnyGoal"},{"kind":"project","url":"https://hazel-heejeong-nam.github.io/anygoal/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.26844.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":38,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":9,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"galilai-group/AnyGoal"},"brief":{"text":"AnyGoal benchmarks whether world models can recombine familiar states and actions to handle shifted goals. Across environments and model families, current models consistently fail on combinations underrepresented in expert trajectories, revealing limits in prediction and planning generalization.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":36,"reason":"trending","score":0.6125},{"paper":{"id":"arxiv:2610.25950","arxivId":"2610.25950","doi":"10.48550/arXiv.2610.25950","title":"AdaExploreBench: Benchmarking LLM Agents for Proactive Exploration and Test-time Adaptation in Evolving Environments","abstract":"Large Language Model (LLM) agents are increasingly deployed in unfamiliar and evolving environments, requiring them to proactively explore, understand, and adapt to environmental dynamics. However, existing benchmarks rarely test whether agents can proactively discover how an environment works and adapt as it evolves. To address this gap, we introduce AdaExploreBench, a benchmark for proactive exploration and test-time adaptation in evolving environments. By instantiating environmental dynamics as hidden rules, AdaExploreBench requires agents to actively acquire evidence and infer the rules. Under implicit rule drift, the latent rule changes without announcement, requiring agents to detect from degraded performance that prior knowledge is stale and re-explore. Experiments on eight leading LLMs show that proactive and adaptive exploration remain challenging: GPT-6 Sol achieves only 61.9% calibrated accuracy, while providing sufficient evidence or removing rule drift improves average performance significantly by close to 10%. These results reveal a critical bottleneck in current LLM agents: they can often reason effectively from informative evidence once it is available, but struggle to actively acquire the right evidence and recognize when previously acquired evidence is no longer valid.","authors":[{"name":"Rui Wang","position":0},{"name":"Zishen Lai","position":1},{"name":"Jiayu Liu","position":2},{"name":"Cheng Qian","authorId":"oa:A5139764426","position":3},{"name":"Jeonghwan Kim","position":4},{"name":"Heng Wang","position":5},{"name":"Xuejun Zhang","position":6},{"name":"Jiateng Liu","position":7},{"name":"Bingxuan Li","position":8},{"name":"Shujin Wu","position":9},{"name":"Qing Zong","authorId":"oa:A5093047682","position":10},{"name":"Tianshi Zheng","authorId":"oa:A5031583104","position":11},{"name":"Chunyang Li","position":12},{"name":"Qihan Lin","authorId":"oa:A5135963722","position":13},{"name":"Chen Li","position":14},{"name":"Dadi Guo","position":15},{"name":"Xiusi Chen","position":16},{"name":"Yangqiu Song","authorId":"oa:A5158711285","position":17}],"categories":["cs.CL","cs.AI"],"primaryCategory":"cs.CL","publishedAt":"2026-10-19T07:41:50Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.25950"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.25950"},{"kind":"code","url":"https://github.com/HKUST-KnowComp/AdaExploreBench","label":"HKUST-KnowComp/AdaExploreBench"},{"kind":"project","url":"https://hkust-knowcomp.github.io/AdaExploreBench/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":32,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":3,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"HKUST-KnowComp/AdaExploreBench"},"brief":{"text":"AdaExploreBench tests whether LLM agents can uncover hidden rules and adapt when those rules change. Across eight models, proactive exploration and recognizing stale knowledge remain challenging, though sufficient evidence or no rule drift improves average performance by close to 10%.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":37,"reason":"trending","score":0.6116},{"paper":{"id":"arxiv:2610.26770","arxivId":"2610.26770","doi":"10.48550/arXiv.2610.26770","title":"HarnessPro: Recursive Harness Self-Improvement with Pareto Optimization for Performance-Cost Trade-offs","abstract":"Harness development plays a crucial role in building capable AI agents. Recent approaches improve harnesses through recursive self-improvement (RSI), but their performance gains often come with higher execution costs. We introduce HarnessPro, a label-free method for evolving harnesses to improve task performance while reducing execution cost. In this setting, candidate harnesses are evaluated using LLM-generated tests without access to ground-truth solutions or official tests. Specialized agents diagnose execution trajectories, plan updates with different performance-cost priorities, and implement them as candidate harnesses. Rather than returning a single harness, HarnessPro uses normalized weighted squared-distance selection to provide cost-focused (lite), balanced (medium), and performance-focused (heavy) harnesses for different deployment budgets. Selected harnesses enter the next iteration, where execution feedback guides further updates. Our experiments cover Terminal-Bench 2, SWE-Bench Pro, and ProgramBench, using GPT-5.5 with Codex and GLM-5.2 with Claude Code. Across these benchmarks using GPT-5.5 with Codex, the heavy mode of HarnessPro achieves an average pass-rate gain of 2.0 percentage points and an average per-task cost reduction of 26.1% relative to RHO. This advantage also holds when using GLM-5.2 with Claude Code. Code is available at https://github.com/pzs19/HarnessPro","authors":[{"name":"Zhuoshi Pan","position":0},{"name":"ZhiYuan Feng","position":1},{"name":"Yukuo Cen","authorId":"oa:A5035985651","position":2},{"name":"Honglin Lin","position":3},{"name":"Junru Lu","position":4},{"name":"Zhongxia Chen","authorId":"oa:A5007798032","position":5},{"name":"Shuo Yang","position":6},{"name":"H. Vicky Zhao","position":7},{"name":"Jie Tang","position":8}],"categories":["cs.LG","cs.CL"],"primaryCategory":"cs.LG","publishedAt":"2026-10-19T17:27:38Z","updatedAt":"2026-10-20T06:00:54Z","version":2,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26770"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26770"},{"kind":"code","url":"https://github.com/pzs19/HarnessPro","label":"pzs19/HarnessPro"},{"kind":"project","url":"https://pzs19.github.io/HarnessPro/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.26770.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":28,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":6,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"pzs19/HarnessPro"},"brief":{"text":"HarnessPro recursively improves AI-agent harnesses and offers cost-focused, balanced, and performance-focused options. Across three benchmarks, its performance-focused mode raised average pass rates by 2.0 percentage points and cut per-task costs by 26.1% versus RHO.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":38,"reason":"trending","score":0.6074},{"paper":{"id":"arxiv:2610.20693","arxivId":"2610.20693","doi":"10.48550/arXiv.2610.20693","title":"Orbia: Do Generated Videos Form a Consistent 3D World?","abstract":"Video generators serve as interactive worlds for exploration and simulation, yet realistic and visually appealing frames do not necessarily depict a consistent 3D world. Existing benchmarks assess selected geometric properties or use vision-language models to judge memory and consistency, leaving systematic 3D evaluation underexplored. We introduce Orbia, the first benchmark dedicated to a systematic and comprehensive evaluation of 3D consistency in interactive world models. Through controlled exploration and revisit, we organize the evaluation into three dimensions: input preservation, generated-content persistence, and 3D self-consistency. For input preservation, registered RGB-D references and annotated anchors support appearance, geometry, and spatial stability measurements. For generated-content persistence, we compare revisits to regions absent from the input image with their first generated observations to test whether models preserve what they create. For 3D self-consistency, cross-view depth and feature correspondence, regional scale agreement, and held-out Gaussian rendering test whether generated views describe the same 3D scene. Minute-long exploration and revisits further test long-term consistency. Our construction and annotation pipeline brings together public datasets, self-captured scenes, and simulation, yielding 720 instances across indoor and outdoor environments and six exploration-and-revisit protocols. Evaluation spans 27 text-, action-, and camera-conditioned models through native-interface adapters. Our Orbia benchmark provides a systematic basis for diagnosing 3D inconsistency and developing more persistent video world models.","authors":[{"name":"Dongyue Lu","authorId":"oa:A5158978696","position":0},{"name":"Guian Fang","authorId":"oa:A5102519401","position":1},{"name":"Hur Lim","authorId":"oa:A5158960584","position":2},{"name":"Yonggang Wu","authorId":"oa:A5158941381","position":3},{"name":"Yanrong Wang","authorId":"oa:A5158978791","position":4},{"name":"Hongshan Chen","authorId":"oa:A5158992955","position":5},{"name":"Shengju Qian","authorId":"oa:A5034240662","position":6},{"name":"Yue Liao","authorId":"oa:A5158984165","position":7},{"name":"Wei Chow","authorId":"oa:A5023862930","position":8},{"name":"Lingdong Kong","authorId":"oa:A5086548564","position":9},{"name":"Tianxin Huang","authorId":"oa:A5158969231","position":10},{"name":"Yu Yang","authorId":"oa:A5158948825","position":11},{"name":"Hongji Yang","authorId":"oa:A5158970236","position":12},{"name":"Junchao Huang","authorId":"oa:A5158961241","position":13},{"name":"Ao Liang","authorId":"oa:A5073574784","position":14},{"name":"Yihao Wang","authorId":"oa:A5158997925","position":15},{"name":"Zihan Wang","authorId":"oa:A5158997927","position":16},{"name":"Rong Li","authorId":"oa:A5158942873","position":17},{"name":"Hanlin Chen","authorId":"oa:A5158946448","position":18},{"name":"Xin Wang","authorId":"oa:A5158948157","position":19},{"name":"Wei Tsang Ooi","authorId":"oa:A5072587271","position":20},{"name":"Mike Zheng Shou","authorId":"oa:A5068937750","position":21},{"name":"Shuicheng Yan","authorId":"oa:A5158977296","position":22}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-14T17:59:53Z","updatedAt":"2026-10-18T11:17:54Z","version":2,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.20693"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.20693"},{"kind":"code","url":"https://github.com/worldbench/orbia","label":"worldbench/orbia"},{"kind":"project","url":"https://orbia-bench.github.io"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:04:20Z"},"signals":{"hfUpvotes":{"value":44,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":48,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"worldbench/orbia"},"brief":{"text":"Orbia benchmarks 27 interactive video models for 3D consistency across input preservation, generated-content persistence, and cross-view agreement. Its 720 test instances and six revisit protocols offer a systematic way to diagnose inconsistencies and guide more persistent world models.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":39,"reason":"trending","score":0.6046},{"paper":{"id":"arxiv:2610.23542","arxivId":"2610.23542","doi":"10.48550/arXiv.2610.23542","title":"Execution Before Memory: Diagnosing Memory Failures in Video World Models","abstract":"A memory test for a video world model should say whether the scene was set, whether the asked change was carried out, and whether the last frame shows the result. Current tests fold these into one number, so a low score is easy to read as forgetting. When the change is not carried out, the hidden state transition cannot appear, so the test is not really a test of memory. We introduce MIST, an evaluation that separates the change from the memory it should leave behind. The model first sets the scene while the change stays hidden, and only later is asked to show it, with leakage before that point kept under control. This separation locates the bottleneck: the model sets the scene but does not carry out the instructed change. The same gap appears in general memory evaluations. On that basis we propose M3-Agent (Make the Missing Move), which recursively revises the details of the input instruction so that the model successfully carries out the requested instruction. Relative to the baseline, it more than triples the MIST reveal score. On the benchmarks, WRBench's score T rises by 26.1% and MBench-T's score U by 52.7%. The method meets the problem from the execution side: it shows what the frozen model can do once the move is stated, and it sets a stricter baseline.","authors":[{"name":"Weiyang Jin","authorId":"oa:A5101252846","position":0},{"name":"Jiwen Yu","authorId":"oa:A5112281651","position":1},{"name":"Sihan Xu","authorId":"oa:A5159067288","position":2},{"name":"Yuwei Niu","authorId":"oa:A5123970532","position":3},{"name":"Xihui Liu","authorId":"oa:A5159048757","position":4}],"categories":["cs.CV","cs.CL"],"primaryCategory":"cs.CV","publishedAt":"2026-10-16T14:37:13Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23542"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23542"},{"kind":"code","url":"https://github.com/HKU-MMLab/MIST","label":"HKU-MMLab/MIST"},{"kind":"project","url":"https://hku-mmlab.github.io/MIST/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.23542.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:04:20Z"},"signals":{"hfUpvotes":{"value":34,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":12,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"HKU-MMLab/MIST"}},"rank":40,"reason":"trending","score":0.6012},{"paper":{"id":"arxiv:2610.29997","arxivId":"2610.29997","doi":"10.48550/arXiv.2610.29997","title":"DONIC: Generalist Dexterous Bimanual Tracking Control from Human Demonstrations","abstract":"Despite recent progress in dexterous reference tracking, existing models focus primarily on single-hand manipulation and a limited range of interactions, leaving the bimanual dexterous behaviors common in real life unaddressed. The scarcity of contact-rich bimanual manipulation data and the increased complexity of coordinating contacts make developing a generalist bimanual tracker particularly challenging. Therefore, we present 1) DONIC, the first generalist bimanual tracker jointly trained across operation categories and human-motion sources, together with 2) BiHOI, a multi-view dataset of 1,575 sequences spanning 229 objects and emphasizing sustained bimanual interaction, where reconstructed interactions are retargeted into robot references. DONIC combines proximal policy optimization with behavioral cloning in a feedback loop that recycles verified successful executions into training demonstrations. An object-centered per-hand reward maintains object-motion feedback while supervising each hand according to its reference contact role. Clip-Centered Advantages remove reference-dependent advantage offsets while preserving within-trajectory differences. Together, these designs enable one single policy to learn coordinated manipulation across heterogeneous references. Without finetuning, a single checkpoint achieves 86.79% success on unseen trajectories of training objects and 57.32% on held-out objects, surpassing the state-of-the-art bimanual tracking baselines by 77.36 and 45.73 percentage points.","authors":[{"name":"Mingze Sun","position":0},{"name":"Ziang Cao","position":1},{"name":"Yuan Fei","position":2},{"name":"Qiankuan Tao","position":3},{"name":"Zihan Wang","position":4},{"name":"Tianqi Liu","position":5},{"name":"Junhao Chen","position":6},{"name":"Haiwen Diao","position":7},{"name":"Fangzhou Hong","position":8},{"name":"Zhaoxi Chen","position":9},{"name":"Ziwei Liu","position":10}],"categories":["cs.RO","cs.CV"],"primaryCategory":"cs.RO","publishedAt":"2026-10-21T16:36:02Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.29997"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.29997"},{"kind":"code","url":"https://github.com/Ropedia/DONIC","label":"Ropedia/DONIC"},{"kind":"project","url":"https://ropedia.github.io/DONIC/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.29997.jpg","label":"thumbnail"}],"signals":{"hfUpvotes":{"value":18,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":4,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"Ropedia/DONIC"},"brief":{"text":"DONIC is a single policy for tracking varied bimanual manipulation demonstrations, trained with a new dataset and a loop that reuses successful executions. Without finetuning, it outperforms prior bimanual tracking methods on unseen trajectories and held-out objects.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":41,"reason":"trending","score":0.5997},{"paper":{"id":"arxiv:2610.26853","arxivId":"2610.26853","doi":"10.48550/arXiv.2610.26853","title":"InfiniDream: Open-Ended Photorealistic 3D Worlds Through Procedural Generation","abstract":"Procedural generation plays a key role in synthetic data creation and game development, offering greater controllability, editability, and diversity than static mesh representations. However, VLM-based code generation often produces relatively simple 3D assets, while sophisticated photorealistic procedural systems require extensive manual development. Existing VLM-guided procedural methods remain largely constrained by predefined factories and APIs, limiting their ability to create novel asset types. We introduce InfiniDream, a VLM-driven framework for open-ended procedural 3D world generation. First, we propose Symbolically Constrained Asset Generation, which leverages existing procedural algorithms to create novel, photorealistic assets while preserving editability and seed-based diversity. Second, we develop an end-to-end text-to-world pipeline that combines semantic scene planning with Deterministic Frustum-Guided Grid Placement to construct large-scale 3D worlds with precise object placement, procedural population, and animation. Extensive evaluations on 100 object and 20 scene generation tasks demonstrate improvements across four perceptual metrics. Diversity analysis, ablation studies, and human preference evaluations further validate the effectiveness of our framework. Project page: https://yichuanh.github.io/infinidream-page/","authors":[{"name":"Yi-Chuan Huang","authorId":"oa:A5061652589","position":0},{"name":"Hao-Jen Chien","authorId":"oa:A5053907081","position":1},{"name":"Zheng-Hui Huang","position":2},{"name":"Yu-Lun Liu","position":3},{"name":"Kaipeng Zhang","position":4},{"name":"Zhixiang Wang","position":5}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T17:59:51Z","updatedAt":"2026-10-20T08:36:13Z","version":2,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26853"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26853"},{"kind":"code","url":"https://github.com/yichuanH/InfiniDream","label":"yichuanH/InfiniDream"},{"kind":"project","url":"https://yichuanh.github.io/infinidream-page/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.26853.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":27,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":33,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"yichuanH/InfiniDream"},"brief":{"text":"InfiniDream uses vision-language models and procedural algorithms to generate editable, photorealistic 3D assets and large worlds from text. Across 100 object and 20 scene tasks, it improves four perceptual metrics over existing methods.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":42,"reason":"trending","score":0.5981},{"paper":{"id":"arxiv:2610.27911","arxivId":"2610.27911","doi":"10.48550/arXiv.2610.27911","title":"MMEnterprise-Bench: Evaluating Enterprise Agents over Evolving Multimodal Records","abstract":"Enterprise agents answer questions over workplace records that evolve through team collaboration. As information is shared, corrected, and revised across documents, chats, and figures, multiple versions of the same information accumulate in the workspace. Reliable question answering therefore requires agents to reason over sufficient evidence, resolve cross-source and temporal conflicts, and recognize when the available evidence is insufficient to establish an answer. To systematically evaluate these capabilities, we introduce MMEnterprise-Bench, a benchmark comprising 1,613 questions grounded in evolving multimodal records from simulated collaboration projects, covering three categories: Evidence Reasoning, Conflict Resolution, and Unanswerability. We evaluate 10 proprietary and open-source multimodal LLMs and find that even the strongest models struggle when evidence contains cross-source or temporal conflicts or spans multiple images. Moreover, increasing the retrieval budget or employing a larger retriever improves evidence recall but does not consistently translate into higher answer accuracy. These findings highlight the persistent challenges of reasoning over evolving workplace records and establish MMEnterprise-Bench as a testbed for developing more reliable enterprise agents. Code and data are available at https://mlvlab.github.io/MMEnterprise-Bench/.","authors":[{"name":"Seohyun Lee","position":0},{"name":"Jongha Kim","position":1},{"name":"Dohwan Ko","position":2},{"name":"Ji Soo Lee","position":3},{"name":"Seoung Choi","position":4},{"name":"Jayun Ku","position":5},{"name":"Hyunwoo J. Kim","position":6}],"categories":["cs.AI"],"primaryCategory":"cs.AI","publishedAt":"2026-10-20T09:25:35Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.27911"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.27911"},{"kind":"code","url":"https://github.com/mlvlab/MMEnterprise-Bench","label":"mlvlab/MMEnterprise-Bench"},{"kind":"project","url":"https://mlvlab.github.io/MMEnterprise-Bench/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.27911.jpg","label":"thumbnail"}],"signals":{"hfUpvotes":{"value":26,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":11,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"mlvlab/MMEnterprise-Bench"},"brief":{"text":"MMEnterprise-Bench tests enterprise agents on 1,613 questions about evolving workplace records. Evaluations show that even strong models struggle with conflicting or image-spanning evidence, and better retrieval does not consistently improve answer accuracy.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":43,"reason":"trending","score":0.5953},{"paper":{"id":"arxiv:2610.27628","arxivId":"2610.27628","doi":"10.48550/arXiv.2610.27628","title":"PAR: Progress-Aligned Recurrence for Length Generalization in State Tracking","abstract":"Looped language models can run more iterations at test time, but additional iterations need not help them solve longer problems. For sequential state tracking, final-answer supervision does not specify how far each iteration should advance, and repeated updates can change the scale of the recurrent state. We introduce \\textbf{Progress-Aligned Recurrence (PAR)}, which trains the prediction at iteration $k$ to match the state after the first $k$ operations and rescales recurrent states to their first-iteration norms. On swap-$S_5$, a Qwen2.5-0.5B model trained on at most 32 moves achieves 95.5\\% mean full-state accuracy at 512 moves, or $16\\times$ the training length, compared with 55.7\\% reported for RLT and 0.8\\% for explicit chain of thought on the same backbone. Controlled $S_5$ ablations show that removing norm anchoring reduces the extrapolation factor from $16\\times$ to below $2.2\\times$, and replacing progress-aligned targets with final-answer supervision reduces it to at most $8\\times$. A learned completion readout selects when to answer without a supplied stopping depth. When intermediate labels are unavailable, prefix self-distillation induces the correspondence from shorter problem prefixes using only final answers. The gains apply across sequential state-tracking tasks, establishing the task structure required for depth extrapolation. These findings support training intermediate predictions and state scale together to make additional recurrent depth useful for longer problems.","authors":[{"name":"Pan Wang","position":0},{"name":"Yihao Hu","position":1},{"name":"Guangyuan Dong","position":2},{"name":"Yitong Lin","position":3},{"name":"Hang Wang","position":4},{"name":"Zimi Liu","position":5},{"name":"Zhihao Wen","position":6},{"name":"Jianshe Li","position":7},{"name":"Jiang-Ming Yang","position":8},{"name":"Wei Wu","position":9},{"name":"Yongqi Tong","position":10}],"categories":["cs.CL"],"primaryCategory":"cs.CL","publishedAt":"2026-10-20T04:54:06Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.27628"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.27628"}],"signals":{"hfUpvotes":{"value":148,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"}},"brief":{"text":"Progress-Aligned Recurrence trains intermediate predictions to track each step and stabilizes recurrent state scale, enabling a Qwen2.5-0.5B model trained on up to 32 swaps to achieve 95.5% full-state accuracy at 512 moves. Ablations show both components matter for extending performance to longer sequences.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":44,"reason":"trending","score":0.5945},{"paper":{"id":"arxiv:2610.26216","arxivId":"2610.26216","doi":"10.48550/arXiv.2610.26216","title":"RoAR-VAE: Improving Latent Diffusability through Robust Prediction","abstract":"An effective visual tokenizer must preserve image fidelity while producing latents that downstream diffusion models can learn efficiently. Prior work improves latent diffusability through semantic alignment or joint tokenizer--diffusion training. Revisiting REPA-E, a representative approach, we find that removing its diffusion objective and dedicated generative blocks while retaining semantic prediction from corrupted latents yields comparable downstream generation performance under a matched setting. This observation motivates a general principle of robust prediction: generation-relevant information should remain recoverable from corrupted latents. We instantiate this principle with two complementary objectives. RobustAlign predicts pretrained semantic features from corrupted latents to preserve high-level structure, while RobustRecon reconstructs the clean image to retain fine-grained spatial and appearance information. Together, they form Robust Alignment and Reconstruction VAE (RoAR-VAE). Our analysis suggests that robust prediction improves diffusability by making useful information more accessible in high-variance latent directions, which diffusion models tend to learn earlier. Empirically, RoAR-VAE improves average generation benchmark performance by 2.19 points over the strongest baseline tokenizer under the same training budget, while improving small-text reconstruction accuracy by 16.93 percentage points over FLUX.1-VAE. These results demonstrate that RoAR-VAE achieves a favorable balance between reconstruction fidelity and downstream generative learnability.","authors":[{"name":"Yang Yue","position":0},{"name":"Fangyun Wei","authorId":"oa:A5090973869","position":1},{"name":"Yitong Wang","position":2},{"name":"Jinjing Zhao","position":3},{"name":"Henry Zheng","authorId":"oa:A5101182803","position":4},{"name":"Dong Chen","position":5},{"name":"Gao Huang","position":6}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T11:42:21Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26216"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26216"},{"kind":"code","url":"https://github.com/LeapLabTHU/RoAR-VAE","label":"LeapLabTHU/RoAR-VAE"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.26216.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":31,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":9,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"LeapLabTHU/RoAR-VAE"},"brief":{"text":"RoAR-VAE trains image latents to retain semantic and fine-detail information even when corrupted, making them easier for diffusion models to learn. It beats the strongest matched-budget tokenizer by 2.19 benchmark points and improves small-text reconstruction accuracy by 16.93 percentage points over FLUX.1-VAE.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":45,"reason":"trending","score":0.591},{"paper":{"id":"arxiv:2610.26067","arxivId":"2610.26067","doi":"10.48550/arXiv.2610.26067","title":"PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning","abstract":"Generalizing real-time physics-based animation across diverse 3D shapes and spatial discretizations remains challenging. In this paper, we introduce PhysSkin++, a physics-informed framework that learns continuous neural skinning fields from static geometries to construct compact deformation subspaces. Specifically, we propose a Neural Skinning Fields Autoencoder that predicts shape-dependent deformation subspaces without per-object network training. To learn these subspaces without simulation trajectories, we develop a Physics-Informed Self-Supervised Learning strategy that combines skinning-field normalization with conflict-aware gradient correction, stabilizing joint optimization under physical and orthogonality constraints. The learned representation supports efficient reduced dynamics with user control and contact interactions, together with direct animation binding to meshes, point clouds, and 3D Gaussians. For prescribed rig animation, we further introduce a self-supervised complementarity objective to learn complementary elastic responses without enforcing complementarity constraints at runtime. Experiments demonstrate cross-shape generalization of the learned skinning fields, real-time reduced simulation, and applications to controllable, collision-aware, and rig-complementary animation.","authors":[{"name":"Yuanhang Lei","position":0},{"name":"Baofei Liu","position":1},{"name":"Xingxuan Li","position":2},{"name":"Tao Cheng","authorId":"oa:A5128691210","position":3},{"name":"Zixuan Wang","position":4},{"name":"Boming Zhao","authorId":"oa:A5103281510","position":5},{"name":"Yujun Shen","position":6},{"name":"Siyuan Huang","position":7},{"name":"Ruizhen Hu","authorId":"oa:A5013892799","position":8},{"name":"Peter Yichen Chen","position":9},{"name":"Hujun Bao","authorId":"oa:A5059021869","position":10},{"name":"Zhaopeng Cui","authorId":"oa:A5075311772","position":11}],"categories":["cs.CV","cs.GR"],"primaryCategory":"cs.CV","publishedAt":"2026-10-19T09:25:55Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.26067"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.26067"},{"kind":"code","url":"https://github.com/zju3dv/PhysSkin","label":"zju3dv/PhysSkin"},{"kind":"project","url":"https://zju3dv.github.io/PhysSkin_plus/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":18,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":35,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"zju3dv/PhysSkin"},"brief":{"text":"PhysSkin++ learns continuous skinning fields from static 3D shapes, enabling real-time physics-based animation across different shapes and mesh resolutions without per-object training. It supports controllable, collision-aware deformation and animation binding to meshes, point clouds, and 3D Gaussians.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":46,"reason":"trending","score":0.5904},{"paper":{"id":"arxiv:2610.23230","arxivId":"2610.23230","doi":"10.48550/arXiv.2610.23230","title":"RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars","abstract":"Making a conference poster is visual coding in miniature: an entire paper must be compressed onto a single page, which exercises summarization, visual design, and code implementation at once, using nothing beyond the paper itself. Yet benchmarks for this task in the context of visual coding remain scarce: existing resources are small and annotated only with layout boxes, so evaluation must fall back on VLM-as-judge or human checklists rather than objective render-and-compare, and the generation methods they support design layouts from scratch, leaving human-crafted posters unusable as either a retrievable design library or as verified supervision. We introduce RAP (Reconstruct A Poster), a dataset and benchmark for paper-to-poster generation as a visual coding task. RAP contains 25K+ verified paper-poster pairs from various machine learning conferences held between 2022 and 2026, annotated with a three-level content taxonomy, plus a held-out benchmark of 150 test papers each paired with its human poster. A fully automated derendering pipeline turns each poster into styled HTML/CSS of typed, refillable slots, so every sample is a paper-poster-markup triplet verifiable by rendering against the source poster. The corpus serves both inference and training: PosterRAG, a training-free multi-agent pipeline, retrieves an exemplar and refills its slots without touching layout or styling, while an SFT+RL recipe trains a dedicated model from the same triplets. An online demo is available at https://reconstruct-a-poster.github.io/.","authors":[{"name":"Haonan Han","authorId":"oa:A5040802446","position":0},{"name":"Yukang Lin","authorId":"oa:A5159088386","position":1},{"name":"Minliang Lin","authorId":"oa:A5159059943","position":2},{"name":"Haoxian Tan","authorId":"oa:A5071239924","position":3},{"name":"Jun-Yan He","authorId":"oa:A5159035830","position":4},{"name":"Rui Yang","authorId":"oa:A5159090684","position":5},{"name":"Zhi-Qi Cheng","authorId":"oa:A5159093086","position":6},{"name":"Jie Hu","authorId":"oa:A5159063537","position":7},{"name":"Guanglu Wan","authorId":"oa:A5072537055","position":8},{"name":"Xiu Li","authorId":"oa:A5159042274","position":9}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-16T09:59:22Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23230"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23230"},{"kind":"code","url":"https://github.com/VincentHancoder/RAP","label":"VincentHancoder/RAP"},{"kind":"project","url":"https://reconstruct-a-poster.github.io/"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.23230.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":36,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":29,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"VincentHancoder/RAP"}},"rank":47,"reason":"trending","score":0.5886},{"paper":{"id":"arxiv:2610.23787","arxivId":"2610.23787","doi":"10.48550/arXiv.2610.23787","title":"DLite: History-free Speculative Decoding for Agentic AI","abstract":"Speculative decoding has become a mainstream technique for accelerating large language model (LLM) inference by using a lightweight draft model to generate candidate tokens that are subsequently verified by the target model. However, agentic workloads introduce long and continuously evolving contexts and frequent multi-turn interactions. Under such workloads, we observe that existing draft models suffer from up to an 80% degradation in acceptance length, while maintaining their historical KV cache introduces substantial system overhead. In this paper, we identify the growing historical context as a common source of both problems and introduce DLite, a history-free speculative decoding framework for agentic AI. DLite directly reuses contextual hidden states from the target model and performs local speculative prediction without explicitly maintaining the historical context, transforming an increasingly complex modeling problem into local simple prediction problem. A lightweight sequential head further restores intra-block causal dependencies with low overhead. To effectively train this heterogeneous history-free architecture, we introduce a tailored two-stage training strategy. Experiments on general-purpose, long-context and agentic workloads show that DLite achieves 2-6 acceptance length and outperforms the state-of-the-art speculative decoding method, DSpark, up to 70%.","authors":[{"name":"Jiaming Xu","authorId":"oa:A5159071579","position":0},{"name":"Hanzhen Wang","authorId":"oa:A5158858416","position":1},{"name":"Tianlang Zhao","authorId":"oa:A5008722002","position":2},{"name":"Jiewen Xiao","authorId":"oa:A5000624228","position":3},{"name":"Chengze Yuan","authorId":"oa:A5159094557","position":4},{"name":"Jiayi Pan","authorId":"oa:A5159067624","position":5},{"name":"Yongkang Zhou","authorId":"oa:A5159083319","position":6},{"name":"Jiancai Ye","authorId":"oa:A5159078078","position":7},{"name":"Yu Wang","authorId":"oa:A5159081425","position":8},{"name":"Guohao Dai","authorId":"oa:A5015946486","position":9}],"categories":["cs.CL"],"primaryCategory":"cs.CL","publishedAt":"2026-10-16T17:18:54Z","version":1,"license":"http://creativecommons.org/licenses/by/4.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23787"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23787"},{"kind":"code","url":"https://github.com/SJTU-DAI-Sys/DLite","label":"SJTU-DAI-Sys/DLite"},{"kind":"other","url":"https://science-os.github.io/scienceos-atlas/thumbs/2610.23787.jpg","label":"thumbnail"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:04:20Z"},"signals":{"hfUpvotes":{"value":33,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":43,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"SJTU-DAI-Sys/DLite"}},"rank":48,"reason":"trending","score":0.5862},{"paper":{"id":"arxiv:2610.23986","arxivId":"2610.23986","doi":"10.48550/arXiv.2610.23986","title":"BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References","abstract":"Recent advances in multimodal large language models (MLLMs) have enabled them to act as autonomous agents, creating and editing 3D models via code or direct software manipulation. In practical workflows, these models must infer object shapes, materials, spatial layouts, and lighting from diverse references while strictly following instructions to reconstruct, modify, or design 3D content. However, existing benchmarks provide limited coverage of such open-ended requests across real-world scenarios. To bridge this gap, we introduce BlenderWorld, the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation tasks. BlenderWorld comprises 223 tasks spanning architecture, products, natural environments, and scientific visualization, driven by text, image, video, and structured specification inputs. To rigorously assess these complex deliverables, we establish a unified evaluation protocol: we render the submitted Blender projects from multiple viewpoints and evaluate them against 2,651 task-specific binary rubrics alongside holistic visual quality scores. Benchmarking thirteen frontier models reveals that GPT-6-Astra achieves the highest overall fused score of 79.9, followed by Opus-5 (73.0) and GPT-5.6-Sol (67.7). Ultimately, our findings reveal a critical bottleneck: while current models can parse broad instructions, they still profoundly struggle with fine-grained 3D spatial coherence and accurate structural reproduction in Blender, leaving substantial headroom for future 3D-native agentics.","authors":[{"name":"Xinyu Fang","position":0},{"name":"Yuchi Wang","position":1},{"name":"Weihao Yu","position":2},{"name":"Rongyao Fang","authorId":"oa:A5025456709","position":3},{"name":"Zhaohai Li","position":4},{"name":"Zhixun Li","position":5},{"name":"Tiancheng Gu","authorId":"oa:A5101333470","position":6},{"name":"Xilin Wei","authorId":"oa:A5113244350","position":7},{"name":"Yuhang Zang","authorId":"oa:A5005200501","position":8},{"name":"Hongsheng Li","position":9},{"name":"Dahua Lin","position":10},{"name":"Guofeng Zhang","position":11},{"name":"Shuai Bai","authorId":"oa:A5101014470","position":12},{"name":"Zhibo Yang","position":13},{"name":"Dayiheng Liu","position":14}],"categories":["cs.CV"],"primaryCategory":"cs.CV","publishedAt":"2026-10-14T12:46:54Z","version":1,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.23986"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.23986"},{"kind":"code","url":"https://github.com/QwenLM/BlenderWorld","label":"QwenLM/BlenderWorld"},{"kind":"project","url":"https://blenderworld.github.io/"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":29,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":187,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"QwenLM/BlenderWorld"},"brief":{"text":"BlenderWorld benchmarks 13 frontier models on 223 real-world 3D creation tasks using varied reference inputs and a unified evaluation protocol. GPT-6-Astra scores highest, but models still struggle with spatial coherence and accurate reconstruction.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":49,"reason":"trending","score":0.5843},{"paper":{"id":"arxiv:2610.24791","arxivId":"2610.24791","doi":"10.48550/arXiv.2610.24791","title":"UniAVC: Unified Audio-Video Creation via Progressive Training and Asymmetric Scheduling","abstract":"Unified audio-video creation requires a single model capable of text-conditioned joint synthesis, bidirectional cross-modal generation, and instruction-guided editing across both modalities. Existing methods rely heavily on paired corpora, which are scarce for joint generation and even rarer for editing. Consequently, they are spent learning single-branch generative and editing capabilities that abundant unimodal data could readily teach, rather than on the cross-modal coupling. Furthermore, prevailing systems apply a single shared noise schedule to both modalities, neglecting two critical limitations. First, due to the significant discrepancy in information density, video requires prolonged high noise to organize spatial and motion structure, whereas the sparser audio signal can settle earlier during the synthesis process, the dynamics a uniform schedule might compromise. Second, lockstep denoising leaves asymmetric workflows, such as conditioning on one modality to synthesize or edit another, untrained. We present UniAVC, which addresses both obstacles. It first trains each branch on unimodal dataset to full generative and editing competence, and subsequently dedicates paired data to exploit cross-modal correspondence. Modality-decoupled schedulers meanwhile warp the noise schedule independently from a shared timestep, while subtasks with one clean modality directly supervise asymmetric configurations. These settings respectively enable intent-aware editing, which preserves one modality intact while enlisting it to constrain the other, and post-inference self-refinement, which regenerates only an unsatisfactory branch against an accepted one. UniAVC thus spans multiple creation tasks from a single set of weights without task-specific adapters.","authors":[{"name":"Kien T. Pham","authorId":"oa:A5132922560","position":0},{"name":"Yuhao Liu","position":1},{"name":"Chong Hou Choi","authorId":"oa:A5133543823","position":2},{"name":"Zhenqi He","authorId":"oa:A5102718144","position":3},{"name":"Yanghao Wang","authorId":"oa:A5101486200","position":4},{"name":"Chenyang Lei","authorId":"oa:A5088135057","position":5},{"name":"Rui Liu","position":6},{"name":"Qifeng Chen","position":7},{"name":"Long Chen","position":8}],"categories":["cs.CV","cs.MM","cs.SD","eess.AS"],"primaryCategory":"cs.CV","publishedAt":"2026-10-17T17:46:17Z","updatedAt":"2026-10-20T07:19:16Z","version":2,"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","links":[{"kind":"abs","url":"https://arxiv.org/abs/2610.24791"},{"kind":"pdf","url":"https://arxiv.org/pdf/2610.24791"},{"kind":"code","url":"https://github.com/HKUST-LongGroup/UniAVC","label":"HKUST-LongGroup/UniAVC"},{"kind":"project","url":"https://hkust-longgroup.github.io/UniAVC"}],"citations":{"value":0,"source":"openalex","fetchedAt":"2026-10-22T08:09:23Z"},"signals":{"hfUpvotes":{"value":10,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubStars":{"value":62,"source":"hf","fetchedAt":"2026-10-22T08:36:18Z"},"githubRepo":"HKUST-LongGroup/UniAVC"},"brief":{"text":"UniAVC trains audio and video branches on abundant single-modality data before using paired data to learn cross-modal links, with separate noise schedules and asymmetric supervision. A single model supports joint generation, cross-modal editing and refinement without task-specific adapters.","source":"llm","model":"gpt-6-luna","generatedAt":"2026-10-22T08:36:18Z"}},"rank":50,"reason":"trending","score":0.58}]}